diff --git a/src/content/book/ar/00a-preface.mdx b/src/content/book/ar/00a-preface.mdx
new file mode 100644
index 00000000..8e84af9e
--- /dev/null
+++ b/src/content/book/ar/00a-preface.mdx
@@ -0,0 +1,82 @@
+
+
+
+
Fatih Kadir Akın
+
Creator of Awesome ChatGPT Prompts (a.k.a. prompts.chat)
+
+ Software developer from Istanbul, leading Developer Relations at Teknasyon. Author of books on JavaScript and prompt engineering. Open-source advocate specializing in web technologies and AI-assisted development.
+
+
+
+
+
+ GPT-4o
+ 128K tokens
+
+
+ GPT-5
+ 400K tokens
+
+
+ Claude Sonnet 4
+ 1M tokens
+
+
+ Gemini 2.5
+ 1M tokens
+
+
+ Llama 4
+ 1M-10M tokens
+
+
+ DeepSeek R1
+ 128K tokens
+
+
+
+ Quick questions
+ Faster models like GPT-4o or Claude 3.5 Sonnet
+
+
+ Hard problems
+ Smarter models like GPT-5.2 or Claude 4.5 Opus
+
+
+ Writing code
+ Code-focused models or the smartest general models
+
+
+ Long documents
+ Models with big context windows (Claude, Gemini)
+
+
+ Current events
+ Models with internet access
+
+
+
+ Creation
+ Write, Create, Generate, Compose, Design
+
+
+ Analysis
+ Analyze, Evaluate, Compare, Assess, Review
+
+
+ Transformation
+ Convert, Translate, Reformat, Summarize, Expand
+
+
+ Explanation
+ Explain, Describe, Clarify, Define, Illustrate
+
+
+ Problem-solving
+ Solve, Debug, Fix, Optimize, Improve
+
+
+
+ Audience
+ Who will read/use this?
+
+
+ Length
+ How long/short should it be?
+
+
+ Tone
+ Formal? Casual? Technical?
+
+
+ Format
+ Prose? List? Table? Code?
+
+
+ Scope
+ What to include/exclude?
+
+
+ Purpose
+ What should this accomplish?
+
+
+
+ Too verbose
+ Add "Be concise" or length limits
+
+
+ Too vague
+ Add specific examples or constraints
+
+
+ Wrong format
+ Specify exact output structure
+
+
+ Missing aspects
+ Add "Make sure to include..."
+
+
+ Wrong tone
+ Specify audience and style
+
+
+ Inaccurate
+ Request citations or step-by-step reasoning
+
+
+
+ Code review
+ Senior developer + mentor
+
+
+ Writing feedback
+ Editor + target audience member
+
+
+ Business strategy
+ Consultant + industry expert
+
+
+ Learning new topic
+ Patient teacher + practitioner
+
+
+ Creative writing
+ Specific genre author
+
+
+ Technical explanation
+ Expert + communicator
+
+
+ Problem-solving
+ Domain specialist + generalist
+
+
+
"Let's think step by step."
+
"Think through this carefully."
+
"Work through this problem methodically."
+
"Show your reasoning at each step."
+
"Explain your thought process."
+
+
+
Use CoT For
+
+
**Math problems** — Reduces calculation errors
+
**Logic puzzles** — Prevents skipped steps
+
**Complex analysis** — Organizes thinking
+
**Code debugging** — Traces execution
+
**Decision making** — Weighs trade-offs
+
+
+
+
Skip CoT For
+
+
**Simple Q&A** — Unnecessary overhead
+
**Creative writing** — Can constrain creativity
+
**Factual lookups** — No reasoning needed
+
**Translation** — Direct task
+
**Summarization** — Usually straightforward
+
+
+
+
+ Simple classification
+ 2-3
+ One per category minimum
+
+
+ Complex formatting
+ 3-5
+ Show variations
+
+
+ Nuanced style
+ 4-6
+ Capture full range
+
+
+ Edge cases
+ 1-2
+ Alongside normal examples
+
+
+
+ Symptom
+ Likely Cause
+ Solution
+
+
+ Too long
+ No length constraint
+ Add word/sentence limits
+
+
+ Too short
+ Lacks detail request
+ Ask for elaboration
+
+
+ Off-topic
+ Vague instructions
+ Be more specific
+
+
+ Wrong format
+ Format not specified
+ Define exact structure
+
+
+ Wrong tone
+ Audience not clear
+ Specify audience/style
+
+
+ Inconsistent
+ No examples provided
+ Add few-shot examples
+
+
+
+
Ready to Ship
+
+
Output consistently meets requirements
+
Edge cases are handled appropriately
+
Format is reliable and parseable
+
Further improvements show diminishing returns
+
+
+
+
Keep Iterating
+
+
Output is inconsistent across runs
+
Edge cases cause failures
+
Critical requirements are missed
+
You haven't tested enough variations
+
+
+
+
+
Use JSON When
+
+
Programmatic parsing needed
+
API responses
+
Strict type requirements
+
JavaScript/Web integration
+
Compact representation
+
+
+
+
Use YAML When
+
+
Human readability matters
+
Configuration files
+
Comments are needed
+
DevOps/Infrastructure
+
Deep nested structures
+
+
+
+
+
Do
+
+ - Give a clear identity
+ - List specific capabilities
+ - Set explicit boundaries
+ - Define the tone and style
+ - Include example responses
+
+
+
+
Don't
+
+ - Be vague about the role
+ - Forget to set limits
+ - Make it too long (500 words max)
+ - Contradict yourself
+ - Assume the AI will "figure it out"
+
+
+
+
+
Single Prompt Struggles
+
+
Multi-step reasoning gets confused
+
Different "modes" of thinking clash
+
Complex outputs lack consistency
+
No opportunity for quality control
+
+
+
+
Chaining Solves This
+
+
Each step focuses on one task
+
Specialized prompts for each mode
+
Validate between steps
+
Debug and improve individual steps
+
+
+
+
+
→
+
+
→
+
+
+
Prompt 3
+
(Generate)
+
+
Output
+
+
+
Best For
+
Document summarization, report generation, content repurposing, data-to-narrative conversion
+
+
Best For
+
Code refactoring, project planning, troubleshooting, strategic decision-making, complex problem solving
+
+
Best For
+
Marketing copy, creative writing, email drafts, presentations, any content that benefits from revision
+
+
+
Reducing Latency
+
+
Parallelize independent steps
+
Cache intermediate results
+
Use smaller models for simple steps
+
Batch similar operations
+
+
+
+
Reducing Cost
+
+
Use cheaper models for classification
+
Limit iterations in loops
+
Short-circuit when possible
+
Cache repeated queries
+
+
+
+
Improving Reliability
+
+
Add validation between steps
+
Include retry logic
+
Log intermediate results
+
Implement fallback paths
+
+
+
+
+
Chaining Enables
+
+
Complex multi-step workflows
+
Higher quality through specialization
+
Better error handling and validation
+
Modular, reusable prompt components
+
+
+
+
Key Principles
+
+
Break complex tasks into simple steps
+
Design clear interfaces between steps
+
Validate intermediate outputs
+
Build in error handling and fallbacks
+
Optimize for your constraints
+
+
+
+
How RAG Works:
+
+
+ 1
+ User asks: "What's our refund policy?"
+
+
+ 2
+ System searches your documents for "refund policy"
+
+
+ 3
+ Finds relevant section from your policy document
+
+
+ 4
+ Sends to AI: "Based on this policy: [text], answer: What's our refund policy?"
+
+
+ 5
+ AI generates accurate answer using your actual policy
+
+
+
+
+
RAG Advantages
+
+ - Uses your actual, current data
+ - Reduces hallucinations
+ - Can cite sources
+ - Easy to update (just update documents)
+ - No expensive fine-tuning needed
+
+
+
+
When to Use RAG
+
+ - Customer support bots
+ - Documentation search
+ - Internal knowledge bases
+ - Any domain-specific Q&A
+ - When accuracy matters
+
+
+
+
+
System Prompts
+
Define the agent's identity, capabilities, constraints, and behavioral guidelines. These are the agent's "constitution."
+
+
+
Planning Prompts
+
Guide how agents break down complex goals into actionable steps. Critical for multi-step reasoning.
+
+
+
Tool-Use Prompts
+
Describe available tools and when/how to use them. Agents must understand their capabilities.
+
+
+
Reflection Prompts
+
Enable agents to evaluate their own outputs, catch errors, and improve iteratively.
+
+
+
+
+
→
+
+
→
+
+
→
+
(repeat)
+
+
+
+
Create Plan
+
Break goal into steps
+
+
↓
+
+
+
→
+
+
→
+
+
→
+
...
+
+
↓
+
+
Revise if Needed
+
Adapt plan based on results
+
+
+
+
+
+
Coordinator
+
Manages workflow
+
+
+
⟷
+
↕
+
+
+
+
+
↓
+
+
Planner Agent
+
Breaks down task
+
+
↓
+
+
Researcher Agent
+
Gathers information
+
+
↓
+
+
Writer Agent
+
Creates content
+
+
↓
+
+
Reviewer Agent
+
Quality checks
+
+
↓
+
+
+
+
+
+ | Today |
+ Tomorrow |
+
+
+
+
+ | Writing detailed prompts |
+ Designing agent systems |
+
+
+ | Manual prompt optimization |
+ Automated prompt tuning |
+
+
+ | Single-model expertise |
+ Multi-model orchestration |
+
+
+ | Text-focused interaction |
+ Multimodal fluency |
+
+
+ | Individual productivity |
+ Team-AI collaboration |
+
+
+
+
+
+
+
+
+
+
Agent
+
Autonomous AI system
+
+
+ powered by
+ ↓
+
+
+
+
+
+
+
Skill
+
Reusable expertise
+
+
+
+
+
+
Skill
+
Reusable expertise
+
+
+
+
+
+
Skill
+
Reusable expertise
+
+
+
+ composed of
+ ↓
+
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
Prompts are atoms → Skills are molecules → Agents are complete structures
+
+
+
+
+
+
→
+
+
+
+ ↺ Loop until complete
+
+
+
→
+
+
+
+
+
+
System Prompts
+
Identity & Role
+
+
+
Planning Prompts
+
How to Think
+
+
+
Tool Prompts
+
How to Act
+
+
+
Recovery Prompts
+
How to Recover
+
+
+
+
📄 SKILL.md (Required)
+
The main instruction file. Contains the core expertise, guidelines, and behaviors that define the skill.
+
+
+
📚 Reference Docs
+
Supporting documentation, examples, and context the agent can reference while working.
+
+
+
🔧 Scripts & Tools
+
Helper scripts, templates, or tool configurations that support the skill's functionality.
+
+
+
⚙️ Configuration
+
Settings, parameters, and customization options for adapting the skill to different contexts.
+
+
+
+ 📁
+ code-review-skill/
+
+
+
+ 📄
+ SKILL.md
+ Core review guidelines
+
+
+ 📄
+ security-checklist.md
+ Security patterns
+
+
+ 📄
+ performance-tips.md
+ Optimization guide
+
+
+
+ 📁
+ language-specific/
+
+
+
+ 📄
+ python.md
+ Python best practices
+
+
+ 📄
+ javascript.md
+ JavaScript patterns
+
+
+ 📄
+ rust.md
+ Rust guidelines
+
+
+
+
+
+
+
Simple Prompt
+
+
Single instruction
+
One-off use
+
Limited context
+
Generic approach
+
No supporting materials
+
+
+
+
Skill
+
+
Comprehensive instruction set
+
Reusable across projects
+
Rich context with references
+
Domain-specific expertise
+
Supporting docs, scripts, configs
+
+
+
+
+ Code Review
+
+
+
+
+ Security Audit
+
+
+
+
+ Documentation
+
+
=
+
+ Complete Code Quality Agent
+
+
+
+ AI Agent
+
+
+
+
Code Review
+
Skill 1
+
+
+
+
Test Writing
+
Skill 3
+
+
+
+
+
+
Core Prompts
+
Planning • Tools • Recovery • Memory
+
+
+
+
+
+
Fatih Kadir Akın
+
Schöpfer von Awesome ChatGPT Prompts (auch bekannt als prompts.chat)
+
+ Softwareentwickler aus Istanbul, leitet Developer Relations bei Teknasyon. Autor von Büchern über JavaScript und Prompt Engineering. Open-Source-Befürworter mit Spezialisierung auf Webtechnologien und KI-gestützte Entwicklung.
+
+
+
+
+
+ GPT-4o
+ 128K tokens
+
+
+ GPT-5
+ 400K tokens
+
+
+ Claude Sonnet 4
+ 1M tokens
+
+
+ Gemini 2.5
+ 1M tokens
+
+
+ Llama 4
+ 1M-10M tokens
+
+
+ DeepSeek R1
+ 128K tokens
+
+
+
+ Schnelle Fragen
+ Schnellere Modelle wie GPT-4o oder Claude 3.5 Sonnet
+
+
+ Schwierige Probleme
+ Intelligentere Modelle wie GPT-5.2 oder Claude 4.5 Opus
+
+
+ Code schreiben
+ Code-fokussierte Modelle oder die intelligentesten allgemeinen Modelle
+
+
+ Lange Dokumente
+ Modelle mit großen Kontextfenstern (Claude, Gemini)
+
+
+ Aktuelle Ereignisse
+ Modelle mit Internetzugang
+
+
+
+ Erstellung
+ Schreibe, Erstelle, Generiere, Verfasse, Entwirf
+
+
+ Analyse
+ Analysiere, Bewerte, Vergleiche, Beurteile, Überprüfe
+
+
+ Transformation
+ Konvertiere, Übersetze, Formatiere um, Fasse zusammen, Erweitere
+
+
+ Erklärung
+ Erkläre, Beschreibe, Kläre, Definiere, Veranschauliche
+
+
+ Problemlösung
+ Löse, Debugge, Behebe, Optimiere, Verbessere
+
+
+
+ Publikum
+ Wer wird das lesen/nutzen?
+
+
+ Länge
+ Wie lang/kurz sollte es sein?
+
+
+ Ton
+ Formell? Locker? Technisch?
+
+
+ Format
+ Prosa? Liste? Tabelle? Code?
+
+
+ Umfang
+ Was einschließen/ausschließen?
+
+
+ Zweck
+ Was soll das erreichen?
+
+
+
+ Zu ausführlich
+ Füge „Sei prägnant" oder Längenlimits hinzu
+
+
+ Zu vage
+ Füge spezifische Beispiele oder Einschränkungen hinzu
+
+
+ Falsches Format
+ Spezifiziere genaue Ausgabestruktur
+
+
+ Fehlende Aspekte
+ Füge „Stelle sicher, dass du einschließt..." hinzu
+
+
+ Falscher Ton
+ Spezifiziere Publikum und Stil
+
+
+ Ungenau
+ Fordere Zitate oder Schritt-für-Schritt-Begründung an
+
+
+
+ Code-Review
+ Senior-Entwickler + Mentor
+
+
+ Schreib-Feedback
+ Lektor + Zielgruppen-Mitglied
+
+
+ Geschäftsstrategie
+ Berater + Branchenexperte
+
+
+ Neues Thema lernen
+ Geduldiger Lehrer + Praktiker
+
+
+ Kreatives Schreiben
+ Spezifischer Genre-Autor
+
+
+ Technische Erklärung
+ Experte + Kommunikator
+
+
+ Problemlösung
+ Domänenspezialist + Generalist
+
+
+
"Lass uns Schritt für Schritt denken."
+
"Denke das sorgfältig durch."
+
"Arbeite dieses Problem methodisch durch."
+
"Zeige deine Begründung bei jedem Schritt."
+
"Erkläre deinen Denkprozess."
+
+
+
CoT verwenden für
+
+
**Mathematische Probleme** — Reduziert Rechenfehler
+
**Logikrätsel** — Verhindert übersprungene Schritte
+
**Komplexe Analysen** — Organisiert das Denken
+
**Code-Debugging** — Verfolgt Ausführung
+
**Entscheidungsfindung** — Wägt Kompromisse ab
+
+
+
+
CoT überspringen für
+
+
**Einfache Fragen** — Unnötiger Overhead
+
**Kreatives Schreiben** — Kann Kreativität einschränken
+
**Faktennachschlagen** — Kein Denken nötig
+
**Übersetzung** — Direkte Aufgabe
+
**Zusammenfassung** — Normalerweise unkompliziert
+
+
+
+
+ Einfache Klassifikation
+ 2-3
+ Mindestens eines pro Kategorie
+
+
+ Komplexe Formatierung
+ 3-5
+ Zeige Variationen
+
+
+ Nuancierter Stil
+ 4-6
+ Erfasse volle Bandbreite
+
+
+ Grenzfälle
+ 1-2
+ Neben normalen Beispielen
+
+
+
+ Symptom
+ Wahrscheinliche Ursache
+ Lösung
+
+
+ Zu lang
+ Keine Längenbeschränkung
+ Wort-/Satzlimits hinzufügen
+
+
+ Zu kurz
+ Fehlende Detailanfrage
+ Um Ausführlichkeit bitten
+
+
+ Thema verfehlt
+ Vage Anweisungen
+ Spezifischer sein
+
+
+ Falsches Format
+ Format nicht spezifiziert
+ Genaue Struktur definieren
+
+
+ Falscher Ton
+ Publikum unklar
+ Publikum/Stil spezifizieren
+
+
+ Inkonsistent
+ Keine Beispiele geliefert
+ Few-Shot-Beispiele hinzufügen
+
+
+
+
Bereit zum Einsatz
+
+
Ausgabe erfüllt konsistent die Anforderungen
+
Grenzfälle werden angemessen behandelt
+
Format ist zuverlässig und parsebar
+
Weitere Verbesserungen zeigen sinkende Erträge
+
+
+
+
Weiter iterieren
+
+
Ausgabe ist inkonsistent über Durchläufe
+
Grenzfälle verursachen Fehler
+
Kritische Anforderungen werden verfehlt
+
Du hast nicht genug Variationen getestet
+
+
+
+
+
JSON verwenden wenn
+
+
Programmatisches Parsing benötigt
+
API-Antworten
+
Strikte Typanforderungen
+
JavaScript/Web-Integration
+
Kompakte Darstellung
+
+
+
+
YAML verwenden wenn
+
+
Menschliche Lesbarkeit wichtig ist
+
Konfigurationsdateien
+
Kommentare benötigt werden
+
DevOps/Infrastruktur
+
Tief verschachtelte Strukturen
+
+
+
+
+
Tu
+
+ - Gib eine klare Identität
+ - Liste spezifische Fähigkeiten auf
+ - Setze explizite Grenzen
+ - Definiere Ton und Stil
+ - Füge Beispielantworten hinzu
+
+
+
+
Tu nicht
+
+ - Sei vage über die Rolle
+ - Vergiss Limits zu setzen
+ - Mach es zu lang (max. 500 Wörter)
+ - Widersprich dir selbst
+ - Geh davon aus, dass die KI es "schon rausfindet"
+
+
+
+
+
Einzelner Prompt kämpft
+
+
Mehrstufiges Denken wird verwirrt
+
Verschiedene Denk-„Modi" kollidieren
+
Komplexen Ausgaben fehlt Konsistenz
+
Keine Möglichkeit zur Qualitätskontrolle
+
+
+
+
Verkettung löst das
+
+
Jeder Schritt fokussiert auf eine Aufgabe
+
Spezialisierte Prompts für jeden Modus
+
Zwischen Schritten validieren
+
Einzelne Schritte debuggen und verbessern
+
+
+
+
+
+
Prompt 1
+
(Extrahieren)
+
+
Eingabe
+
+
→
+
+
+
Prompt 2
+
(Analysieren)
+
+
Zwischenergebnis
+
+
→
+
+
+
Prompt 3
+
(Generieren)
+
+
Ausgabe
+
+
+
Am besten für
+
Dokumentzusammenfassung, Berichtserstellung, Inhaltswiederverwendung, Daten-zu-Narrativ-Konvertierung
+
+
Am besten für
+
Code-Refactoring, Projektplanung, Fehlerbehebung, strategische Entscheidungsfindung, komplexe Problemlösung
+
+
Am besten für
+
Marketingtexte, kreatives Schreiben, E-Mail-Entwürfe, Präsentationen, alle Inhalte, die von Überarbeitung profitieren
+
+
+
Latenz reduzieren
+
+
Unabhängige Schritte parallelisieren
+
Zwischenergebnisse cachen
+
Kleinere Modelle für einfache Schritte
+
Ähnliche Operationen bündeln
+
+
+
+
Kosten reduzieren
+
+
Günstigere Modelle für Klassifikation
+
Iterationen in Schleifen begrenzen
+
Wenn möglich abkürzen
+
Wiederholte Anfragen cachen
+
+
+
+
Zuverlässigkeit verbessern
+
+
Validierung zwischen Schritten
+
Retry-Logik einbauen
+
Zwischenergebnisse loggen
+
Fallback-Pfade implementieren
+
+
+
+
+
Verkettung ermöglicht
+
+
Komplexe mehrstufige Workflows
+
Höhere Qualität durch Spezialisierung
+
Bessere Fehlerbehandlung und Validierung
+
Modulare, wiederverwendbare Prompt-Komponenten
+
+
+
+
Schlüsselprinzipien
+
+
Komplexe Aufgaben in einfache Schritte zerlegen
+
Klare Schnittstellen zwischen Schritten designen
+
Zwischenergebnisse validieren
+
Fehlerbehandlung und Fallbacks einbauen
+
Für deine Einschränkungen optimieren
+
+
+
+
Wie RAG funktioniert:
+
+
+ 1
+ Benutzer fragt: „Was ist unsere Rückgaberichtlinie?"
+
+
+ 2
+ System durchsucht deine Dokumente nach „Rückgaberichtlinie"
+
+
+ 3
+ Findet relevanten Abschnitt aus deinem Richtlinien-Dokument
+
+
+ 4
+ Sendet an KI: „Basierend auf dieser Richtlinie: [Text], beantworte: Was ist unsere Rückgaberichtlinie?"
+
+
+ 5
+ KI generiert genaue Antwort mit deiner tatsächlichen Richtlinie
+
+
+
+
+
RAG-Vorteile
+
+ - Verwendet deine tatsächlichen, aktuellen Daten
+ - Reduziert Halluzinationen
+ - Kann Quellen zitieren
+ - Einfach zu aktualisieren (einfach Dokumente aktualisieren)
+ - Kein teures Fine-Tuning nötig
+
+
+
+
Wann RAG verwenden
+
+ - Kundensupport-Bots
+ - Dokumentationssuche
+ - Interne Wissensdatenbanken
+ - Jede domänenspezifische Q&A
+ - Wenn Genauigkeit wichtig ist
+
+
+
+
+
System Prompts
+
Definieren die Identität, Fähigkeiten, Einschränkungen und Verhaltensrichtlinien des Agents. Diese sind die „Verfassung" des Agents.
+
+
+
Planungs-Prompts
+
Leiten, wie Agents komplexe Ziele in umsetzbare Schritte aufbrechen. Kritisch für mehrstufiges Reasoning.
+
+
+
Tool-Nutzungs-Prompts
+
Beschreiben verfügbare Tools und wann/wie sie zu verwenden sind. Agents müssen ihre Fähigkeiten verstehen.
+
+
+
Reflexions-Prompts
+
Ermöglichen Agents, ihre eigenen Outputs zu evaluieren, Fehler zu erkennen und iterativ zu verbessern.
+
+
+
+
+
→
+
+
→
+
+
→
+
(wiederholen)
+
+
+
+
Plan erstellen
+
Ziel in Schritte aufteilen
+
+
↓
+
+
+
→
+
+
→
+
+
→
+
...
+
+
↓
+
+
Falls nötig überarbeiten
+
Plan basierend auf Ergebnissen anpassen
+
+
+
+
+
+
Koordinator
+
Verwaltet Workflow
+
+
+
⟷
+
↕
+
+
+
+
+
↓
+
+
Planer-Agent
+
Zerlegt Aufgabe
+
+
↓
+
+
Recherche-Agent
+
Sammelt Informationen
+
+
↓
+
+
Schreib-Agent
+
Erstellt Inhalte
+
+
↓
+
+
Reviewer-Agent
+
Qualitätsprüfung
+
+
↓
+
+
+
+
+
+ | Heute |
+ Morgen |
+
+
+
+
+ | Detaillierte Prompts schreiben |
+ Agent-Systeme designen |
+
+
+ | Manuelle Prompt-Optimierung |
+ Automatisiertes Prompt-Tuning |
+
+
+ | Einzelmodell-Expertise |
+ Multi-Modell-Orchestrierung |
+
+
+ | Text-fokussierte Interaktion |
+ Multimodale Gewandtheit |
+
+
+ | Individuelle Produktivität |
+ Team-KI-Zusammenarbeit |
+
+
+
+
+
+
+
+
+
+
Agent
+
Autonomes KI-System
+
+
+ angetrieben von
+ ↓
+
+
+
+
+
+
+
Skill
+
Wiederverwendbare Expertise
+
+
+
+
+
+
Skill
+
Wiederverwendbare Expertise
+
+
+
+
+
+
Skill
+
Wiederverwendbare Expertise
+
+
+
+ zusammengesetzt aus
+ ↓
+
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
Prompts sind Atome → Skills sind Moleküle → Agents sind vollständige Strukturen
+
+
+
+
+
+
→
+
+
+
+ ↺ Schleife bis fertig
+
+
+
→
+
+
+
+
+
+
System Prompts
+
Identität & Rolle
+
+
+
Planungs-Prompts
+
Wie denken
+
+
+
Tool-Prompts
+
Wie handeln
+
+
+
Recovery-Prompts
+
Wie erholen
+
+
+
+
📄 SKILL.md (Erforderlich)
+
Die Haupt-Anweisungsdatei. Enthält die Kern-Expertise, Richtlinien und Verhaltensweisen, die den Skill definieren.
+
+
+
📚 Referenzdokumente
+
Unterstützende Dokumentation, Beispiele und Kontext, die der Agent während der Arbeit referenzieren kann.
+
+
+
🔧 Skripte & Tools
+
Hilfs-Skripte, Vorlagen oder Tool-Konfigurationen, die die Funktionalität des Skills unterstützen.
+
+
+
⚙️ Konfiguration
+
Einstellungen, Parameter und Anpassungsoptionen für die Adaption des Skills an verschiedene Kontexte.
+
+
+
+ 📁
+ code-review-skill/
+
+
+
+ 📄
+ SKILL.md
+ Kern-Review-Richtlinien
+
+
+ 📄
+ security-checklist.md
+ Sicherheitsmuster
+
+
+ 📄
+ performance-tips.md
+ Optimierungsleitfaden
+
+
+
+ 📁
+ language-specific/
+
+
+
+ 📄
+ python.md
+ Python Best Practices
+
+
+ 📄
+ javascript.md
+ JavaScript-Muster
+
+
+ 📄
+ rust.md
+ Rust-Richtlinien
+
+
+
+
+
+
+
Einfacher Prompt
+
+
Einzelne Anweisung
+
Einmalige Nutzung
+
Begrenzter Kontext
+
Generischer Ansatz
+
Keine unterstützenden Materialien
+
+
+
+
Skill
+
+
Umfassendes Anweisungsset
+
Wiederverwendbar über Projekte
+
Reicher Kontext mit Referenzen
+
Domänenspezifische Expertise
+
Unterstützende Docs, Skripte, Configs
+
+
+
+
+ Code Review
+
+
+
+
+ Security Audit
+
+
+
+
+ Dokumentation
+
+
=
+
+ Kompletter Code-Qualitäts-Agent
+
+
+
+ KI-Agent
+
+
+
+
Code Review
+
Skill 1
+
+
+
+
Test-Schreiben
+
Skill 3
+
+
+
+
+
+
Kern-Prompts
+
Planung • Tools • Recovery • Gedächtnis
+
+
+
+
+
+
Fatih Kadir Akın
+
Creator of Awesome ChatGPT Prompts (a.k.a. prompts.chat)
+
+ Software developer from Istanbul, leading Developer Relations at Teknasyon. Author of books on JavaScript and prompt engineering. Open-source advocate specializing in web technologies and AI-assisted development.
+
+
+
+
+
+ GPT-4o
+ 128K tokens
+
+
+ GPT-5
+ 400K tokens
+
+
+ Claude Sonnet 4
+ 1M tokens
+
+
+ Gemini 2.5
+ 1M tokens
+
+
+ Llama 4
+ 1M-10M tokens
+
+
+ DeepSeek R1
+ 128K tokens
+
+
+
+ Quick questions
+ Faster models like GPT-4o or Claude 3.5 Sonnet
+
+
+ Hard problems
+ Smarter models like GPT-5.2 or Claude 4.5 Opus
+
+
+ Writing code
+ Code-focused models or the smartest general models
+
+
+ Long documents
+ Models with big context windows (Claude, Gemini)
+
+
+ Current events
+ Models with internet access
+
+
+
+ Creation
+ Write, Create, Generate, Compose, Design
+
+
+ Analysis
+ Analyze, Evaluate, Compare, Assess, Review
+
+
+ Transformation
+ Convert, Translate, Reformat, Summarize, Expand
+
+
+ Explanation
+ Explain, Describe, Clarify, Define, Illustrate
+
+
+ Problem-solving
+ Solve, Debug, Fix, Optimize, Improve
+
+
+
+ Audience
+ Who will read/use this?
+
+
+ Length
+ How long/short should it be?
+
+
+ Tone
+ Formal? Casual? Technical?
+
+
+ Format
+ Prose? List? Table? Code?
+
+
+ Scope
+ What to include/exclude?
+
+
+ Purpose
+ What should this accomplish?
+
+
+
+ Too verbose
+ Add "Be concise" or length limits
+
+
+ Too vague
+ Add specific examples or constraints
+
+
+ Wrong format
+ Specify exact output structure
+
+
+ Missing aspects
+ Add "Make sure to include..."
+
+
+ Wrong tone
+ Specify audience and style
+
+
+ Inaccurate
+ Request citations or step-by-step reasoning
+
+
+
+ Code review
+ Senior developer + mentor
+
+
+ Writing feedback
+ Editor + target audience member
+
+
+ Business strategy
+ Consultant + industry expert
+
+
+ Learning new topic
+ Patient teacher + practitioner
+
+
+ Creative writing
+ Specific genre author
+
+
+ Technical explanation
+ Expert + communicator
+
+
+ Problem-solving
+ Domain specialist + generalist
+
+
+
"Let's think step by step."
+
"Think through this carefully."
+
"Work through this problem methodically."
+
"Show your reasoning at each step."
+
"Explain your thought process."
+
+
+
Use CoT For
+
+
**Math problems** — Reduces calculation errors
+
**Logic puzzles** — Prevents skipped steps
+
**Complex analysis** — Organizes thinking
+
**Code debugging** — Traces execution
+
**Decision making** — Weighs trade-offs
+
+
+
+
Skip CoT For
+
+
**Simple Q&A** — Unnecessary overhead
+
**Creative writing** — Can constrain creativity
+
**Factual lookups** — No reasoning needed
+
**Translation** — Direct task
+
**Summarization** — Usually straightforward
+
+
+
+
+ Simple classification
+ 2-3
+ One per category minimum
+
+
+ Complex formatting
+ 3-5
+ Show variations
+
+
+ Nuanced style
+ 4-6
+ Capture full range
+
+
+ Edge cases
+ 1-2
+ Alongside normal examples
+
+
+
+ Symptom
+ Likely Cause
+ Solution
+
+
+ Too long
+ No length constraint
+ Add word/sentence limits
+
+
+ Too short
+ Lacks detail request
+ Ask for elaboration
+
+
+ Off-topic
+ Vague instructions
+ Be more specific
+
+
+ Wrong format
+ Format not specified
+ Define exact structure
+
+
+ Wrong tone
+ Audience not clear
+ Specify audience/style
+
+
+ Inconsistent
+ No examples provided
+ Add few-shot examples
+
+
+
+
Ready to Ship
+
+
Output consistently meets requirements
+
Edge cases are handled appropriately
+
Format is reliable and parseable
+
Further improvements show diminishing returns
+
+
+
+
Keep Iterating
+
+
Output is inconsistent across runs
+
Edge cases cause failures
+
Critical requirements are missed
+
You haven't tested enough variations
+
+
+
+
+
Use JSON When
+
+
Programmatic parsing needed
+
API responses
+
Strict type requirements
+
JavaScript/Web integration
+
Compact representation
+
+
+
+
Use YAML When
+
+
Human readability matters
+
Configuration files
+
Comments are needed
+
DevOps/Infrastructure
+
Deep nested structures
+
+
+
+
+
Do
+
+ - Give a clear identity
+ - List specific capabilities
+ - Set explicit boundaries
+ - Define the tone and style
+ - Include example responses
+
+
+
+
Don't
+
+ - Be vague about the role
+ - Forget to set limits
+ - Make it too long (500 words max)
+ - Contradict yourself
+ - Assume the AI will "figure it out"
+
+
+
+
+
Single Prompt Struggles
+
+
Multi-step reasoning gets confused
+
Different "modes" of thinking clash
+
Complex outputs lack consistency
+
No opportunity for quality control
+
+
+
+
Chaining Solves This
+
+
Each step focuses on one task
+
Specialized prompts for each mode
+
Validate between steps
+
Debug and improve individual steps
+
+
+
+
+
→
+
+
→
+
+
+
Prompt 3
+
(Generate)
+
+
Output
+
+
+
Best For
+
Document summarization, report generation, content repurposing, data-to-narrative conversion
+
+
Best For
+
Code refactoring, project planning, troubleshooting, strategic decision-making, complex problem solving
+
+
Best For
+
Marketing copy, creative writing, email drafts, presentations, any content that benefits from revision
+
+
+
Reducing Latency
+
+
Parallelize independent steps
+
Cache intermediate results
+
Use smaller models for simple steps
+
Batch similar operations
+
+
+
+
Reducing Cost
+
+
Use cheaper models for classification
+
Limit iterations in loops
+
Short-circuit when possible
+
Cache repeated queries
+
+
+
+
Improving Reliability
+
+
Add validation between steps
+
Include retry logic
+
Log intermediate results
+
Implement fallback paths
+
+
+
+
+
Chaining Enables
+
+
Complex multi-step workflows
+
Higher quality through specialization
+
Better error handling and validation
+
Modular, reusable prompt components
+
+
+
+
Key Principles
+
+
Break complex tasks into simple steps
+
Design clear interfaces between steps
+
Validate intermediate outputs
+
Build in error handling and fallbacks
+
Optimize for your constraints
+
+
+
+
How RAG Works:
+
+
+ 1
+ User asks: "What's our refund policy?"
+
+
+ 2
+ System searches your documents for "refund policy"
+
+
+ 3
+ Finds relevant section from your policy document
+
+
+ 4
+ Sends to AI: "Based on this policy: [text], answer: What's our refund policy?"
+
+
+ 5
+ AI generates accurate answer using your actual policy
+
+
+
+
+
RAG Advantages
+
+ - Uses your actual, current data
+ - Reduces hallucinations
+ - Can cite sources
+ - Easy to update (just update documents)
+ - No expensive fine-tuning needed
+
+
+
+
When to Use RAG
+
+ - Customer support bots
+ - Documentation search
+ - Internal knowledge bases
+ - Any domain-specific Q&A
+ - When accuracy matters
+
+
+
+
+
System Prompts
+
Define the agent's identity, capabilities, constraints, and behavioral guidelines. These are the agent's "constitution."
+
+
+
Planning Prompts
+
Guide how agents break down complex goals into actionable steps. Critical for multi-step reasoning.
+
+
+
Tool-Use Prompts
+
Describe available tools and when/how to use them. Agents must understand their capabilities.
+
+
+
Reflection Prompts
+
Enable agents to evaluate their own outputs, catch errors, and improve iteratively.
+
+
+
+
+
→
+
+
→
+
+
→
+
(repeat)
+
+
+
+
Create Plan
+
Break goal into steps
+
+
↓
+
+
+
→
+
+
→
+
+
→
+
...
+
+
↓
+
+
Revise if Needed
+
Adapt plan based on results
+
+
+
+
+
+
Coordinator
+
Manages workflow
+
+
+
⟷
+
↕
+
+
+
+
+
↓
+
+
Planner Agent
+
Breaks down task
+
+
↓
+
+
Researcher Agent
+
Gathers information
+
+
↓
+
+
Writer Agent
+
Creates content
+
+
↓
+
+
Reviewer Agent
+
Quality checks
+
+
↓
+
+
+
+
+
+ | Today |
+ Tomorrow |
+
+
+
+
+ | Writing detailed prompts |
+ Designing agent systems |
+
+
+ | Manual prompt optimization |
+ Automated prompt tuning |
+
+
+ | Single-model expertise |
+ Multi-model orchestration |
+
+
+ | Text-focused interaction |
+ Multimodal fluency |
+
+
+ | Individual productivity |
+ Team-AI collaboration |
+
+
+
+
+
+
+
+
+
+
Agent
+
Autonomous AI system
+
+
+ powered by
+ ↓
+
+
+
+
+
+
+
Skill
+
Reusable expertise
+
+
+
+
+
+
Skill
+
Reusable expertise
+
+
+
+
+
+
Skill
+
Reusable expertise
+
+
+
+ composed of
+ ↓
+
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
Prompts are atoms → Skills are molecules → Agents are complete structures
+
+
+
+
+
+
→
+
+
+
+ ↺ Loop until complete
+
+
+
→
+
+
+
+
+
+
System Prompts
+
Identity & Role
+
+
+
Planning Prompts
+
How to Think
+
+
+
Tool Prompts
+
How to Act
+
+
+
Recovery Prompts
+
How to Recover
+
+
+
+
📄 SKILL.md (Required)
+
The main instruction file. Contains the core expertise, guidelines, and behaviors that define the skill.
+
+
+
📚 Reference Docs
+
Supporting documentation, examples, and context the agent can reference while working.
+
+
+
🔧 Scripts & Tools
+
Helper scripts, templates, or tool configurations that support the skill's functionality.
+
+
+
⚙️ Configuration
+
Settings, parameters, and customization options for adapting the skill to different contexts.
+
+
+
+ 📁
+ code-review-skill/
+
+
+
+ 📄
+ SKILL.md
+ Core review guidelines
+
+
+ 📄
+ security-checklist.md
+ Security patterns
+
+
+ 📄
+ performance-tips.md
+ Optimization guide
+
+
+
+ 📁
+ language-specific/
+
+
+
+ 📄
+ python.md
+ Python best practices
+
+
+ 📄
+ javascript.md
+ JavaScript patterns
+
+
+ 📄
+ rust.md
+ Rust guidelines
+
+
+
+
+
+
+
Simple Prompt
+
+
Single instruction
+
One-off use
+
Limited context
+
Generic approach
+
No supporting materials
+
+
+
+
Skill
+
+
Comprehensive instruction set
+
Reusable across projects
+
Rich context with references
+
Domain-specific expertise
+
Supporting docs, scripts, configs
+
+
+
+
+ Code Review
+
+
+
+
+ Security Audit
+
+
+
+
+ Documentation
+
+
=
+
+ Complete Code Quality Agent
+
+
+
+ AI Agent
+
+
+
+
Code Review
+
Skill 1
+
+
+
+
Test Writing
+
Skill 3
+
+
+
+
+
+
Core Prompts
+
Planning • Tools • Recovery • Memory
+
+
+
+
+
+
Fatih Kadir Akın
+
Creador de Awesome ChatGPT Prompts (también conocido como prompts.chat)
+
+ Desarrollador de software de Estambul, liderando Relaciones con Desarrolladores en Teknasyon. Autor de libros sobre JavaScript e ingeniería de prompts. Defensor del código abierto especializado en tecnologías web y desarrollo asistido por IA.
+
+
+
+
+
+ GPT-4o
+ 128K tokens
+
+
+ GPT-5
+ 400K tokens
+
+
+ Claude Sonnet 4
+ 1M tokens
+
+
+ Gemini 2.5
+ 1M tokens
+
+
+ Llama 4
+ 1M-10M tokens
+
+
+ DeepSeek R1
+ 128K tokens
+
+
+
+ Preguntas rápidas
+ Modelos más rápidos como GPT-4o o Claude 3.5 Sonnet
+
+
+ Problemas difíciles
+ Modelos más inteligentes como GPT-5.2 o Claude 4.5 Opus
+
+
+ Escribir código
+ Modelos enfocados en código o los modelos generales más inteligentes
+
+
+ Documentos largos
+ Modelos con grandes ventanas de contexto (Claude, Gemini)
+
+
+ Eventos actuales
+ Modelos con acceso a internet
+
+
+
+ Creación
+ Escribe, Crea, Genera, Compone, Diseña
+
+
+ Análisis
+ Analiza, Evalúa, Compara, Valora, Revisa
+
+
+ Transformación
+ Convierte, Traduce, Reformatea, Resume, Expande
+
+
+ Explicación
+ Explica, Describe, Aclara, Define, Ilustra
+
+
+ Solución de problemas
+ Resuelve, Depura, Arregla, Optimiza, Mejora
+
+
+
+ Audiencia
+ ¿Quién leerá/usará esto?
+
+
+ Longitud
+ ¿Qué tan largo/corto debe ser?
+
+
+ Tono
+ ¿Formal? ¿Casual? ¿Técnico?
+
+
+ Formato
+ ¿Prosa? ¿Lista? ¿Tabla? ¿Código?
+
+
+ Alcance
+ ¿Qué incluir/excluir?
+
+
+ Propósito
+ ¿Qué debe lograr esto?
+
+
+
+ Muy verboso
+ Agrega "Sé conciso" o límites de longitud
+
+
+ Muy vago
+ Agrega ejemplos específicos o restricciones
+
+
+ Formato incorrecto
+ Especifica la estructura de salida exacta
+
+
+ Aspectos faltantes
+ Agrega "Asegúrate de incluir..."
+
+
+ Tono incorrecto
+ Especifica audiencia y estilo
+
+
+ Inexacto
+ Solicita citas o razonamiento paso a paso
+
+
+
+ Revisión de código
+ Desarrollador senior + mentor
+
+
+ Retroalimentación de escritura
+ Editor + miembro de audiencia objetivo
+
+
+ Estrategia de negocios
+ Consultor + experto de industria
+
+
+ Aprender tema nuevo
+ Maestro paciente + practicante
+
+
+ Escritura creativa
+ Autor de género específico
+
+
+ Explicación técnica
+ Experto + comunicador
+
+
+ Resolución de problemas
+ Especialista de dominio + generalista
+
+
+
"Pensemos paso a paso."
+
"Piensa esto cuidadosamente."
+
"Trabaja este problema metódicamente."
+
"Muestra tu razonamiento en cada paso."
+
"Explica tu proceso de pensamiento."
+
+
+
Usa CoT Para
+
+
**Problemas matemáticos** — Reduce errores de cálculo
+
**Puzzles de lógica** — Previene pasos saltados
+
**Análisis complejos** — Organiza el pensamiento
+
**Depuración de código** — Rastrea la ejecución
+
**Toma de decisiones** — Sopesa compensaciones
+
+
+
+
Omite CoT Para
+
+
**Preguntas y respuestas simples** — Overhead innecesario
+
**Escritura creativa** — Puede restringir la creatividad
+
**Búsquedas factuales** — No se necesita razonamiento
+
**Traducción** — Tarea directa
+
**Resumen** — Generalmente directo
+
+
+
+
+ Clasificación simple
+ 2-3
+ Uno por categoría mínimo
+
+
+ Formateo complejo
+ 3-5
+ Muestra variaciones
+
+
+ Estilo matizado
+ 4-6
+ Captura el rango completo
+
+
+ Casos límite
+ 1-2
+ Junto con ejemplos normales
+
+
+
+ Síntoma
+ Causa Probable
+ Solución
+
+
+ Muy largo
+ Sin restricción de longitud
+ Agregar límites de palabras/oraciones
+
+
+ Muy corto
+ Falta solicitud de detalle
+ Pedir elaboración
+
+
+ Fuera de tema
+ Instrucciones vagas
+ Ser más específico
+
+
+ Formato incorrecto
+ Formato no especificado
+ Definir estructura exacta
+
+
+ Tono incorrecto
+ Audiencia no clara
+ Especificar audiencia/estilo
+
+
+ Inconsistente
+ Sin ejemplos proporcionados
+ Agregar ejemplos de pocos ejemplos
+
+
+
+
Listo para Enviar
+
+
La salida cumple consistentemente los requisitos
+
Los casos límite se manejan apropiadamente
+
El formato es confiable y parseable
+
Mejoras adicionales muestran rendimientos decrecientes
+
+
+
+
Sigue Iterando
+
+
La salida es inconsistente entre ejecuciones
+
Los casos límite causan fallos
+
Se pierden requisitos críticos
+
No has probado suficientes variaciones
+
+
+
+
+
Usa JSON Cuando
+
+
Se necesita parseo programático
+
Respuestas de API
+
Requisitos estrictos de tipos
+
Integración JavaScript/Web
+
Representación compacta
+
+
+
+
Usa YAML Cuando
+
+
Importa la legibilidad humana
+
Archivos de configuración
+
Se necesitan comentarios
+
DevOps/Infraestructura
+
Estructuras profundamente anidadas
+
+
+
+
+
Hacer
+
+ - Dar una identidad clara
+ - Listar capacidades específicas
+ - Establecer límites explícitos
+ - Definir el tono y estilo
+ - Incluir respuestas de ejemplo
+
+
+
+
No Hacer
+
+ - Ser vago sobre el rol
+ - Olvidar establecer límites
+ - Hacerlo muy largo (500 palabras máx)
+ - Contradecirte a ti mismo
+ - Asumir que la IA "lo descubrirá"
+
+
+
+
+
Problemas del Prompt Único
+
+
El razonamiento de múltiples pasos se confunde
+
Diferentes "modos" de pensamiento chocan
+
Las salidas complejas carecen de consistencia
+
Sin oportunidad para control de calidad
+
+
+
+
El Encadenamiento Resuelve Esto
+
+
Cada paso se enfoca en una tarea
+
Prompts especializados para cada modo
+
Validar entre pasos
+
Depurar y mejorar pasos individuales
+
+
+
+
+
→
+
+
+
Prompt 2
+
(Analizar)
+
+
Intermedio
+
+
→
+
+
+
Mejor Para
+
Resumen de documentos, generación de informes, reutilización de contenido, conversión de datos a narrativa
+
+
Mejor Para
+
Refactorización de código, planificación de proyectos, resolución de problemas, toma de decisiones estratégicas, resolución de problemas complejos
+
+
Mejor Para
+
Copy de marketing, escritura creativa, borradores de email, presentaciones, cualquier contenido que se beneficie de revisión
+
+
+
Reduciendo Latencia
+
+
Paraleliza pasos independientes
+
Cachea resultados intermedios
+
Usa modelos más pequeños para pasos simples
+
Agrupa operaciones similares
+
+
+
+
Reduciendo Costo
+
+
Usa modelos más baratos para clasificación
+
Limita iteraciones en bucles
+
Cortocircuita cuando sea posible
+
Cachea consultas repetidas
+
+
+
+
Mejorando Confiabilidad
+
+
Agrega validación entre pasos
+
Incluye lógica de reintentos
+
Registra resultados intermedios
+
Implementa rutas de respaldo
+
+
+
+
+
El Encadenamiento Habilita
+
+
Flujos de trabajo complejos de múltiples pasos
+
Mayor calidad a través de especialización
+
Mejor manejo de errores y validación
+
Componentes de prompts modulares y reutilizables
+
+
+
+
Principios Clave
+
+
Divide tareas complejas en pasos simples
+
Diseña interfaces claras entre pasos
+
Valida salidas intermedias
+
Incorpora manejo de errores y respaldos
+
Optimiza para tus restricciones
+
+
+
+
Cómo Funciona RAG:
+
+
+ 1
+ Usuario pregunta: "¿Cuál es nuestra política de reembolso?"
+
+
+ 2
+ El sistema busca en tus documentos "política de reembolso"
+
+
+ 3
+ Encuentra la sección relevante de tu documento de políticas
+
+
+ 4
+ Envía a la IA: "Basándote en esta política: [texto], responde: ¿Cuál es nuestra política de reembolso?"
+
+
+ 5
+ La IA genera respuesta precisa usando tu política real
+
+
+
+
+
Ventajas de RAG
+
+ - Usa tus datos reales y actuales
+ - Reduce alucinaciones
+ - Puede citar fuentes
+ - Fácil de actualizar (solo actualiza documentos)
+ - No necesita fine-tuning costoso
+
+
+
+
Cuándo Usar RAG
+
+ - Bots de soporte al cliente
+ - Búsqueda de documentación
+ - Bases de conocimiento internas
+ - Cualquier Q&A específico de dominio
+ - Cuando la precisión importa
+
+
+
+
+
Prompts de Sistema
+
Definen la identidad, capacidades, restricciones y directrices de comportamiento del agente. Estos son la "constitución" del agente.
+
+
+
Prompts de Planificación
+
Guían cómo los agentes descomponen objetivos complejos en pasos accionables. Críticos para razonamiento multi-paso.
+
+
+
Prompts de Uso de Herramientas
+
Describen herramientas disponibles y cuándo/cómo usarlas. Los agentes deben entender sus capacidades.
+
+
+
Prompts de Reflexión
+
Permiten a los agentes evaluar sus propias salidas, detectar errores, y mejorar iterativamente.
+
+
+
+
+
→
+
+
→
+
+
→
+
(repetir)
+
+
+
+
Crear Plan
+
Dividir objetivo en pasos
+
+
↓
+
+
+
→
+
+
→
+
+
→
+
...
+
+
↓
+
+
Revisar si es Necesario
+
Adaptar plan basado en resultados
+
+
+
+
+
+
Coordinador
+
Gestiona flujo de trabajo
+
+
+
⟷
+
↕
+
+
+
+
+
Solicitud del Usuario
+
+
↓
+
+
Agente Planificador
+
Descompone la tarea
+
+
↓
+
+
Agente Investigador
+
Recopila información
+
+
↓
+
+
Agente Escritor
+
Crea contenido
+
+
↓
+
+
Agente Revisor
+
Verificación de calidad
+
+
↓
+
+
+
+
+
+ | Hoy |
+ Mañana |
+
+
+
+
+ | Escribir prompts detallados |
+ Diseñar sistemas de agentes |
+
+
+ | Optimización manual de prompts |
+ Ajuste automatizado de prompts |
+
+
+ | Experiencia en un solo modelo |
+ Orquestación multi-modelo |
+
+
+ | Interacción enfocada en texto |
+ Fluidez multimodal |
+
+
+ | Productividad individual |
+ Colaboración equipo-IA |
+
+
+
+
+
+
+
+
+
+
Agente
+
Sistema de IA autónomo
+
+
+ impulsado por
+ ↓
+
+
+
+
+
+
+
Skill
+
Experiencia reutilizable
+
+
+
+
+
+
Skill
+
Experiencia reutilizable
+
+
+
+
+
+
Skill
+
Experiencia reutilizable
+
+
+
+ compuesto de
+ ↓
+
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
Los Prompts son átomos → Los Skills son moléculas → Los Agentes son estructuras completas
+
+
+
+
+
+
→
+
+
+
+ ↺ Repetir hasta completar
+
+
+
→
+
+
+
+
+
+
Prompts de Sistema
+
Identidad y Rol
+
+
+
Prompts de Planificación
+
Cómo Pensar
+
+
+
Prompts de Herramientas
+
Cómo Actuar
+
+
+
Prompts de Recuperación
+
Cómo Recuperarse
+
+
+
+
📄 SKILL.md (Requerido)
+
El archivo de instrucciones principal. Contiene la experiencia central, directrices y comportamientos que definen el skill.
+
+
+
📚 Documentación de Referencia
+
Documentación de apoyo, ejemplos y contexto que el agente puede referenciar mientras trabaja.
+
+
+
🔧 Scripts y Herramientas
+
Scripts auxiliares, plantillas o configuraciones de herramientas que apoyan la funcionalidad del skill.
+
+
+
⚙️ Configuración
+
Ajustes, parámetros y opciones de personalización para adaptar el skill a diferentes contextos.
+
+
+
+ 📁
+ code-review-skill/
+
+
+
+ 📄
+ SKILL.md
+ Directrices de revisión
+
+
+ 📄
+ security-checklist.md
+ Patrones de seguridad
+
+
+ 📄
+ performance-tips.md
+ Guía de optimización
+
+
+
+ 📁
+ language-specific/
+
+
+
+ 📄
+ python.md
+ Mejores prácticas Python
+
+
+ 📄
+ javascript.md
+ Patrones JavaScript
+
+
+ 📄
+ rust.md
+ Directrices Rust
+
+
+
+
+
+
+
Prompt Simple
+
+
Instrucción única
+
Uso de una sola vez
+
Contexto limitado
+
Enfoque genérico
+
Sin materiales de apoyo
+
+
+
+
Skill
+
+
Conjunto completo de instrucciones
+
Reutilizable entre proyectos
+
Contexto rico con referencias
+
Experiencia específica de dominio
+
Docs, scripts, configs de apoyo
+
+
+
+
+ Revisión de Código
+
+
+
+
+ Auditoría de Seguridad
+
+
+
+
+ Documentación
+
+
=
+
+ Agente de Calidad de Código Completo
+
+
+
+ Agente de IA
+
+
+
+
Revisión de Código
+
Skill 1
+
+
+
Diseño de API
+
Skill 2
+
+
+
Escritura de Tests
+
Skill 3
+
+
+
+
+
+
Prompts Centrales
+
Planificación • Herramientas • Recuperación • Memoria
+
+
+
+
+
+
Fatih Kadir Akın
+
Creator of Awesome ChatGPT Prompts (a.k.a. prompts.chat)
+
+ Software developer from Istanbul, leading Developer Relations at Teknasyon. Author of books on JavaScript and prompt engineering. Open-source advocate specializing in web technologies and AI-assisted development.
+
+
+
+
+
+ GPT-4o
+ 128K tokens
+
+
+ GPT-5
+ 400K tokens
+
+
+ Claude Sonnet 4
+ 1M tokens
+
+
+ Gemini 2.5
+ 1M tokens
+
+
+ Llama 4
+ 1M-10M tokens
+
+
+ DeepSeek R1
+ 128K tokens
+
+
+
+ Quick questions
+ Faster models like GPT-4o or Claude 3.5 Sonnet
+
+
+ Hard problems
+ Smarter models like GPT-5.2 or Claude 4.5 Opus
+
+
+ Writing code
+ Code-focused models or the smartest general models
+
+
+ Long documents
+ Models with big context windows (Claude, Gemini)
+
+
+ Current events
+ Models with internet access
+
+
+
+ Creation
+ Write, Create, Generate, Compose, Design
+
+
+ Analysis
+ Analyze, Evaluate, Compare, Assess, Review
+
+
+ Transformation
+ Convert, Translate, Reformat, Summarize, Expand
+
+
+ Explanation
+ Explain, Describe, Clarify, Define, Illustrate
+
+
+ Problem-solving
+ Solve, Debug, Fix, Optimize, Improve
+
+
+
+ Audience
+ Who will read/use this?
+
+
+ Length
+ How long/short should it be?
+
+
+ Tone
+ Formal? Casual? Technical?
+
+
+ Format
+ Prose? List? Table? Code?
+
+
+ Scope
+ What to include/exclude?
+
+
+ Purpose
+ What should this accomplish?
+
+
+
+ Too verbose
+ Add "Be concise" or length limits
+
+
+ Too vague
+ Add specific examples or constraints
+
+
+ Wrong format
+ Specify exact output structure
+
+
+ Missing aspects
+ Add "Make sure to include..."
+
+
+ Wrong tone
+ Specify audience and style
+
+
+ Inaccurate
+ Request citations or step-by-step reasoning
+
+
+
+ Code review
+ Senior developer + mentor
+
+
+ Writing feedback
+ Editor + target audience member
+
+
+ Business strategy
+ Consultant + industry expert
+
+
+ Learning new topic
+ Patient teacher + practitioner
+
+
+ Creative writing
+ Specific genre author
+
+
+ Technical explanation
+ Expert + communicator
+
+
+ Problem-solving
+ Domain specialist + generalist
+
+
+
"Let's think step by step."
+
"Think through this carefully."
+
"Work through this problem methodically."
+
"Show your reasoning at each step."
+
"Explain your thought process."
+
+
+
Use CoT For
+
+
**Math problems** — Reduces calculation errors
+
**Logic puzzles** — Prevents skipped steps
+
**Complex analysis** — Organizes thinking
+
**Code debugging** — Traces execution
+
**Decision making** — Weighs trade-offs
+
+
+
+
Skip CoT For
+
+
**Simple Q&A** — Unnecessary overhead
+
**Creative writing** — Can constrain creativity
+
**Factual lookups** — No reasoning needed
+
**Translation** — Direct task
+
**Summarization** — Usually straightforward
+
+
+
+
+ Simple classification
+ 2-3
+ One per category minimum
+
+
+ Complex formatting
+ 3-5
+ Show variations
+
+
+ Nuanced style
+ 4-6
+ Capture full range
+
+
+ Edge cases
+ 1-2
+ Alongside normal examples
+
+
+
+ Symptom
+ Likely Cause
+ Solution
+
+
+ Too long
+ No length constraint
+ Add word/sentence limits
+
+
+ Too short
+ Lacks detail request
+ Ask for elaboration
+
+
+ Off-topic
+ Vague instructions
+ Be more specific
+
+
+ Wrong format
+ Format not specified
+ Define exact structure
+
+
+ Wrong tone
+ Audience not clear
+ Specify audience/style
+
+
+ Inconsistent
+ No examples provided
+ Add few-shot examples
+
+
+
+
Ready to Ship
+
+
Output consistently meets requirements
+
Edge cases are handled appropriately
+
Format is reliable and parseable
+
Further improvements show diminishing returns
+
+
+
+
Keep Iterating
+
+
Output is inconsistent across runs
+
Edge cases cause failures
+
Critical requirements are missed
+
You haven't tested enough variations
+
+
+
+
+
Use JSON When
+
+
Programmatic parsing needed
+
API responses
+
Strict type requirements
+
JavaScript/Web integration
+
Compact representation
+
+
+
+
Use YAML When
+
+
Human readability matters
+
Configuration files
+
Comments are needed
+
DevOps/Infrastructure
+
Deep nested structures
+
+
+
+
+
Do
+
+ - Give a clear identity
+ - List specific capabilities
+ - Set explicit boundaries
+ - Define the tone and style
+ - Include example responses
+
+
+
+
Don't
+
+ - Be vague about the role
+ - Forget to set limits
+ - Make it too long (500 words max)
+ - Contradict yourself
+ - Assume the AI will "figure it out"
+
+
+
+
+
Single Prompt Struggles
+
+
Multi-step reasoning gets confused
+
Different "modes" of thinking clash
+
Complex outputs lack consistency
+
No opportunity for quality control
+
+
+
+
Chaining Solves This
+
+
Each step focuses on one task
+
Specialized prompts for each mode
+
Validate between steps
+
Debug and improve individual steps
+
+
+
+
+
→
+
+
→
+
+
+
Prompt 3
+
(Generate)
+
+
Output
+
+
+
Best For
+
Document summarization, report generation, content repurposing, data-to-narrative conversion
+
+
Best For
+
Code refactoring, project planning, troubleshooting, strategic decision-making, complex problem solving
+
+
Best For
+
Marketing copy, creative writing, email drafts, presentations, any content that benefits from revision
+
+
+
Reducing Latency
+
+
Parallelize independent steps
+
Cache intermediate results
+
Use smaller models for simple steps
+
Batch similar operations
+
+
+
+
Reducing Cost
+
+
Use cheaper models for classification
+
Limit iterations in loops
+
Short-circuit when possible
+
Cache repeated queries
+
+
+
+
Improving Reliability
+
+
Add validation between steps
+
Include retry logic
+
Log intermediate results
+
Implement fallback paths
+
+
+
+
+
Chaining Enables
+
+
Complex multi-step workflows
+
Higher quality through specialization
+
Better error handling and validation
+
Modular, reusable prompt components
+
+
+
+
Key Principles
+
+
Break complex tasks into simple steps
+
Design clear interfaces between steps
+
Validate intermediate outputs
+
Build in error handling and fallbacks
+
Optimize for your constraints
+
+
+
+
How RAG Works:
+
+
+ 1
+ User asks: "What's our refund policy?"
+
+
+ 2
+ System searches your documents for "refund policy"
+
+
+ 3
+ Finds relevant section from your policy document
+
+
+ 4
+ Sends to AI: "Based on this policy: [text], answer: What's our refund policy?"
+
+
+ 5
+ AI generates accurate answer using your actual policy
+
+
+
+
+
RAG Advantages
+
+ - Uses your actual, current data
+ - Reduces hallucinations
+ - Can cite sources
+ - Easy to update (just update documents)
+ - No expensive fine-tuning needed
+
+
+
+
When to Use RAG
+
+ - Customer support bots
+ - Documentation search
+ - Internal knowledge bases
+ - Any domain-specific Q&A
+ - When accuracy matters
+
+
+
+
+
System Prompts
+
Define the agent's identity, capabilities, constraints, and behavioral guidelines. These are the agent's "constitution."
+
+
+
Planning Prompts
+
Guide how agents break down complex goals into actionable steps. Critical for multi-step reasoning.
+
+
+
Tool-Use Prompts
+
Describe available tools and when/how to use them. Agents must understand their capabilities.
+
+
+
Reflection Prompts
+
Enable agents to evaluate their own outputs, catch errors, and improve iteratively.
+
+
+
+
+
→
+
+
→
+
+
→
+
(repeat)
+
+
+
+
Create Plan
+
Break goal into steps
+
+
↓
+
+
+
→
+
+
→
+
+
→
+
...
+
+
↓
+
+
Revise if Needed
+
Adapt plan based on results
+
+
+
+
+
+
Coordinator
+
Manages workflow
+
+
+
⟷
+
↕
+
+
+
+
+
↓
+
+
Planner Agent
+
Breaks down task
+
+
↓
+
+
Researcher Agent
+
Gathers information
+
+
↓
+
+
Writer Agent
+
Creates content
+
+
↓
+
+
Reviewer Agent
+
Quality checks
+
+
↓
+
+
+
+
+
+ | Today |
+ Tomorrow |
+
+
+
+
+ | Writing detailed prompts |
+ Designing agent systems |
+
+
+ | Manual prompt optimization |
+ Automated prompt tuning |
+
+
+ | Single-model expertise |
+ Multi-model orchestration |
+
+
+ | Text-focused interaction |
+ Multimodal fluency |
+
+
+ | Individual productivity |
+ Team-AI collaboration |
+
+
+
+
+
+
+
+
+
+
Agent
+
Autonomous AI system
+
+
+ powered by
+ ↓
+
+
+
+
+
+
+
Skill
+
Reusable expertise
+
+
+
+
+
+
Skill
+
Reusable expertise
+
+
+
+
+
+
Skill
+
Reusable expertise
+
+
+
+ composed of
+ ↓
+
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
Prompts are atoms → Skills are molecules → Agents are complete structures
+
+
+
+
+
+
→
+
+
+
+ ↺ Loop until complete
+
+
+
→
+
+
+
+
+
+
System Prompts
+
Identity & Role
+
+
+
Planning Prompts
+
How to Think
+
+
+
Tool Prompts
+
How to Act
+
+
+
Recovery Prompts
+
How to Recover
+
+
+
+
📄 SKILL.md (Required)
+
The main instruction file. Contains the core expertise, guidelines, and behaviors that define the skill.
+
+
+
📚 Reference Docs
+
Supporting documentation, examples, and context the agent can reference while working.
+
+
+
🔧 Scripts & Tools
+
Helper scripts, templates, or tool configurations that support the skill's functionality.
+
+
+
⚙️ Configuration
+
Settings, parameters, and customization options for adapting the skill to different contexts.
+
+
+
+ 📁
+ code-review-skill/
+
+
+
+ 📄
+ SKILL.md
+ Core review guidelines
+
+
+ 📄
+ security-checklist.md
+ Security patterns
+
+
+ 📄
+ performance-tips.md
+ Optimization guide
+
+
+
+ 📁
+ language-specific/
+
+
+
+ 📄
+ python.md
+ Python best practices
+
+
+ 📄
+ javascript.md
+ JavaScript patterns
+
+
+ 📄
+ rust.md
+ Rust guidelines
+
+
+
+
+
+
+
Simple Prompt
+
+
Single instruction
+
One-off use
+
Limited context
+
Generic approach
+
No supporting materials
+
+
+
+
Skill
+
+
Comprehensive instruction set
+
Reusable across projects
+
Rich context with references
+
Domain-specific expertise
+
Supporting docs, scripts, configs
+
+
+
+
+ Code Review
+
+
+
+
+ Security Audit
+
+
+
+
+ Documentation
+
+
=
+
+ Complete Code Quality Agent
+
+
+
+ AI Agent
+
+
+
+
Code Review
+
Skill 1
+
+
+
+
Test Writing
+
Skill 3
+
+
+
+
+
+
Core Prompts
+
Planning • Tools • Recovery • Memory
+
+
+
+
+
+
Fatih Kadir Akın
+
Creator of Awesome ChatGPT Prompts (a.k.a. prompts.chat)
+
+ Software developer from Istanbul, leading Developer Relations at Teknasyon. Author of books on JavaScript and prompt engineering. Open-source advocate specializing in web technologies and AI-assisted development.
+
+
+
+
+
+ GPT-4o
+ 128K tokens
+
+
+ GPT-5
+ 400K tokens
+
+
+ Claude Sonnet 4
+ 1M tokens
+
+
+ Gemini 2.5
+ 1M tokens
+
+
+ Llama 4
+ 1M-10M tokens
+
+
+ DeepSeek R1
+ 128K tokens
+
+
+
+ Quick questions
+ Faster models like GPT-4o or Claude 3.5 Sonnet
+
+
+ Hard problems
+ Smarter models like GPT-5.2 or Claude 4.5 Opus
+
+
+ Writing code
+ Code-focused models or the smartest general models
+
+
+ Long documents
+ Models with big context windows (Claude, Gemini)
+
+
+ Current events
+ Models with internet access
+
+
+
+ Creation
+ Write, Create, Generate, Compose, Design
+
+
+ Analysis
+ Analyze, Evaluate, Compare, Assess, Review
+
+
+ Transformation
+ Convert, Translate, Reformat, Summarize, Expand
+
+
+ Explanation
+ Explain, Describe, Clarify, Define, Illustrate
+
+
+ Problem-solving
+ Solve, Debug, Fix, Optimize, Improve
+
+
+
+ Audience
+ Who will read/use this?
+
+
+ Length
+ How long/short should it be?
+
+
+ Tone
+ Formal? Casual? Technical?
+
+
+ Format
+ Prose? List? Table? Code?
+
+
+ Scope
+ What to include/exclude?
+
+
+ Purpose
+ What should this accomplish?
+
+
+
+ Too verbose
+ Add "Be concise" or length limits
+
+
+ Too vague
+ Add specific examples or constraints
+
+
+ Wrong format
+ Specify exact output structure
+
+
+ Missing aspects
+ Add "Make sure to include..."
+
+
+ Wrong tone
+ Specify audience and style
+
+
+ Inaccurate
+ Request citations or step-by-step reasoning
+
+
+
+ Code review
+ Senior developer + mentor
+
+
+ Writing feedback
+ Editor + target audience member
+
+
+ Business strategy
+ Consultant + industry expert
+
+
+ Learning new topic
+ Patient teacher + practitioner
+
+
+ Creative writing
+ Specific genre author
+
+
+ Technical explanation
+ Expert + communicator
+
+
+ Problem-solving
+ Domain specialist + generalist
+
+
+
"Let's think step by step."
+
"Think through this carefully."
+
"Work through this problem methodically."
+
"Show your reasoning at each step."
+
"Explain your thought process."
+
+
+
Use CoT For
+
+
**Math problems** — Reduces calculation errors
+
**Logic puzzles** — Prevents skipped steps
+
**Complex analysis** — Organizes thinking
+
**Code debugging** — Traces execution
+
**Decision making** — Weighs trade-offs
+
+
+
+
Skip CoT For
+
+
**Simple Q&A** — Unnecessary overhead
+
**Creative writing** — Can constrain creativity
+
**Factual lookups** — No reasoning needed
+
**Translation** — Direct task
+
**Summarization** — Usually straightforward
+
+
+
+
+ Simple classification
+ 2-3
+ One per category minimum
+
+
+ Complex formatting
+ 3-5
+ Show variations
+
+
+ Nuanced style
+ 4-6
+ Capture full range
+
+
+ Edge cases
+ 1-2
+ Alongside normal examples
+
+
+
+ Symptom
+ Likely Cause
+ Solution
+
+
+ Too long
+ No length constraint
+ Add word/sentence limits
+
+
+ Too short
+ Lacks detail request
+ Ask for elaboration
+
+
+ Off-topic
+ Vague instructions
+ Be more specific
+
+
+ Wrong format
+ Format not specified
+ Define exact structure
+
+
+ Wrong tone
+ Audience not clear
+ Specify audience/style
+
+
+ Inconsistent
+ No examples provided
+ Add few-shot examples
+
+
+
+
Ready to Ship
+
+
Output consistently meets requirements
+
Edge cases are handled appropriately
+
Format is reliable and parseable
+
Further improvements show diminishing returns
+
+
+
+
Keep Iterating
+
+
Output is inconsistent across runs
+
Edge cases cause failures
+
Critical requirements are missed
+
You haven't tested enough variations
+
+
+
+
+
Use JSON When
+
+
Programmatic parsing needed
+
API responses
+
Strict type requirements
+
JavaScript/Web integration
+
Compact representation
+
+
+
+
Use YAML When
+
+
Human readability matters
+
Configuration files
+
Comments are needed
+
DevOps/Infrastructure
+
Deep nested structures
+
+
+
+
+
Do
+
+ - Give a clear identity
+ - List specific capabilities
+ - Set explicit boundaries
+ - Define the tone and style
+ - Include example responses
+
+
+
+
Don't
+
+ - Be vague about the role
+ - Forget to set limits
+ - Make it too long (500 words max)
+ - Contradict yourself
+ - Assume the AI will "figure it out"
+
+
+
+
+
Single Prompt Struggles
+
+
Multi-step reasoning gets confused
+
Different "modes" of thinking clash
+
Complex outputs lack consistency
+
No opportunity for quality control
+
+
+
+
Chaining Solves This
+
+
Each step focuses on one task
+
Specialized prompts for each mode
+
Validate between steps
+
Debug and improve individual steps
+
+
+
+
+
→
+
+
→
+
+
+
Prompt 3
+
(Generate)
+
+
Output
+
+
+
Best For
+
Document summarization, report generation, content repurposing, data-to-narrative conversion
+
+
Best For
+
Code refactoring, project planning, troubleshooting, strategic decision-making, complex problem solving
+
+
Best For
+
Marketing copy, creative writing, email drafts, presentations, any content that benefits from revision
+
+
+
Reducing Latency
+
+
Parallelize independent steps
+
Cache intermediate results
+
Use smaller models for simple steps
+
Batch similar operations
+
+
+
+
Reducing Cost
+
+
Use cheaper models for classification
+
Limit iterations in loops
+
Short-circuit when possible
+
Cache repeated queries
+
+
+
+
Improving Reliability
+
+
Add validation between steps
+
Include retry logic
+
Log intermediate results
+
Implement fallback paths
+
+
+
+
+
Chaining Enables
+
+
Complex multi-step workflows
+
Higher quality through specialization
+
Better error handling and validation
+
Modular, reusable prompt components
+
+
+
+
Key Principles
+
+
Break complex tasks into simple steps
+
Design clear interfaces between steps
+
Validate intermediate outputs
+
Build in error handling and fallbacks
+
Optimize for your constraints
+
+
+
+
How RAG Works:
+
+
+ 1
+ User asks: "What's our refund policy?"
+
+
+ 2
+ System searches your documents for "refund policy"
+
+
+ 3
+ Finds relevant section from your policy document
+
+
+ 4
+ Sends to AI: "Based on this policy: [text], answer: What's our refund policy?"
+
+
+ 5
+ AI generates accurate answer using your actual policy
+
+
+
+
+
RAG Advantages
+
+ - Uses your actual, current data
+ - Reduces hallucinations
+ - Can cite sources
+ - Easy to update (just update documents)
+ - No expensive fine-tuning needed
+
+
+
+
When to Use RAG
+
+ - Customer support bots
+ - Documentation search
+ - Internal knowledge bases
+ - Any domain-specific Q&A
+ - When accuracy matters
+
+
+
+
+
System Prompts
+
Define the agent's identity, capabilities, constraints, and behavioral guidelines. These are the agent's "constitution."
+
+
+
Planning Prompts
+
Guide how agents break down complex goals into actionable steps. Critical for multi-step reasoning.
+
+
+
Tool-Use Prompts
+
Describe available tools and when/how to use them. Agents must understand their capabilities.
+
+
+
Reflection Prompts
+
Enable agents to evaluate their own outputs, catch errors, and improve iteratively.
+
+
+
+
+
→
+
+
→
+
+
→
+
(repeat)
+
+
+
+
Create Plan
+
Break goal into steps
+
+
↓
+
+
+
→
+
+
→
+
+
→
+
...
+
+
↓
+
+
Revise if Needed
+
Adapt plan based on results
+
+
+
+
+
+
Coordinator
+
Manages workflow
+
+
+
⟷
+
↕
+
+
+
+
+
↓
+
+
Planner Agent
+
Breaks down task
+
+
↓
+
+
Researcher Agent
+
Gathers information
+
+
↓
+
+
Writer Agent
+
Creates content
+
+
↓
+
+
Reviewer Agent
+
Quality checks
+
+
↓
+
+
+
+
+
+ | Today |
+ Tomorrow |
+
+
+
+
+ | Writing detailed prompts |
+ Designing agent systems |
+
+
+ | Manual prompt optimization |
+ Automated prompt tuning |
+
+
+ | Single-model expertise |
+ Multi-model orchestration |
+
+
+ | Text-focused interaction |
+ Multimodal fluency |
+
+
+ | Individual productivity |
+ Team-AI collaboration |
+
+
+
+
+
+
+
+
+
+
Agent
+
Autonomous AI system
+
+
+ powered by
+ ↓
+
+
+
+
+
+
+
Skill
+
Reusable expertise
+
+
+
+
+
+
Skill
+
Reusable expertise
+
+
+
+
+
+
Skill
+
Reusable expertise
+
+
+
+ composed of
+ ↓
+
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
Prompts are atoms → Skills are molecules → Agents are complete structures
+
+
+
+
+
+
→
+
+
+
+ ↺ Loop until complete
+
+
+
→
+
+
+
+
+
+
System Prompts
+
Identity & Role
+
+
+
Planning Prompts
+
How to Think
+
+
+
Tool Prompts
+
How to Act
+
+
+
Recovery Prompts
+
How to Recover
+
+
+
+
📄 SKILL.md (Required)
+
The main instruction file. Contains the core expertise, guidelines, and behaviors that define the skill.
+
+
+
📚 Reference Docs
+
Supporting documentation, examples, and context the agent can reference while working.
+
+
+
🔧 Scripts & Tools
+
Helper scripts, templates, or tool configurations that support the skill's functionality.
+
+
+
⚙️ Configuration
+
Settings, parameters, and customization options for adapting the skill to different contexts.
+
+
+
+ 📁
+ code-review-skill/
+
+
+
+ 📄
+ SKILL.md
+ Core review guidelines
+
+
+ 📄
+ security-checklist.md
+ Security patterns
+
+
+ 📄
+ performance-tips.md
+ Optimization guide
+
+
+
+ 📁
+ language-specific/
+
+
+
+ 📄
+ python.md
+ Python best practices
+
+
+ 📄
+ javascript.md
+ JavaScript patterns
+
+
+ 📄
+ rust.md
+ Rust guidelines
+
+
+
+
+
+
+
Simple Prompt
+
+
Single instruction
+
One-off use
+
Limited context
+
Generic approach
+
No supporting materials
+
+
+
+
Skill
+
+
Comprehensive instruction set
+
Reusable across projects
+
Rich context with references
+
Domain-specific expertise
+
Supporting docs, scripts, configs
+
+
+
+
+ Code Review
+
+
+
+
+ Security Audit
+
+
+
+
+ Documentation
+
+
=
+
+ Complete Code Quality Agent
+
+
+
+ AI Agent
+
+
+
+
Code Review
+
Skill 1
+
+
+
+
Test Writing
+
Skill 3
+
+
+
+
+
+
Core Prompts
+
Planning • Tools • Recovery • Memory
+
+
+
+
+
+
Fatih Kadir Akın
+
Creatore di Awesome ChatGPT Prompts (noto anche come prompts.chat)
+
+ Sviluppatore software di Istanbul, a capo delle Developer Relations presso Teknasyon. Autore di libri su JavaScript e prompt engineering. Sostenitore dell'open-source specializzato in tecnologie web e sviluppo assistito dall'IA.
+
+
+
+
+
+ GPT-4o
+ 128K token
+
+
+ GPT-5
+ 400K token
+
+
+ Claude Sonnet 4
+ 1M token
+
+
+ Gemini 2.5
+ 1M token
+
+
+ Llama 4
+ 1M-10M token
+
+
+ DeepSeek R1
+ 128K token
+
+
+
+ Domande veloci
+ Modelli più veloci come GPT-4o o Claude 3.5 Sonnet
+
+
+ Problemi difficili
+ Modelli più intelligenti come GPT-5.2 o Claude 4.5 Opus
+
+
+ Scrivere codice
+ Modelli focalizzati sul codice o i modelli generali più intelligenti
+
+
+ Documenti lunghi
+ Modelli con grandi finestre di contesto (Claude, Gemini)
+
+
+ Eventi attuali
+ Modelli con accesso a internet
+
+
+
+ Creazione
+ Scrivi, Crea, Genera, Componi, Progetta
+
+
+ Analisi
+ Analizza, Valuta, Confronta, Esamina, Rivedi
+
+
+ Trasformazione
+ Converti, Traduci, Riformatta, Riassumi, Espandi
+
+
+ Spiegazione
+ Spiega, Descrivi, Chiarisci, Definisci, Illustra
+
+
+ Risoluzione Problemi
+ Risolvi, Debug, Correggi, Ottimizza, Migliora
+
+
+
+ Pubblico
+ Chi leggerà/userà questo?
+
+
+ Lunghezza
+ Quanto lungo/corto dovrebbe essere?
+
+
+ Tono
+ Formale? Informale? Tecnico?
+
+
+ Formato
+ Prosa? Lista? Tabella? Codice?
+
+
+ Ambito
+ Cosa includere/escludere?
+
+
+ Scopo
+ Cosa dovrebbe realizzare questo?
+
+
+
+ Troppo prolisso
+ Aggiungi "Sii conciso" o limiti di lunghezza
+
+
+ Troppo vago
+ Aggiungi esempi specifici o vincoli
+
+
+ Formato sbagliato
+ Specifica la struttura esatta dell'output
+
+
+ Aspetti mancanti
+ Aggiungi "Assicurati di includere..."
+
+
+ Tono sbagliato
+ Specifica pubblico e stile
+
+
+ Impreciso
+ Richiedi citazioni o ragionamento passo-passo
+
+
+
+ Code review
+ Sviluppatore senior + mentore
+
+
+ Feedback scrittura
+ Editor + membro del pubblico target
+
+
+ Strategia business
+ Consulente + esperto del settore
+
+
+ Imparare nuovo argomento
+ Insegnante paziente + praticante
+
+
+ Scrittura creativa
+ Autore di genere specifico
+
+
+ Spiegazione tecnica
+ Esperto + comunicatore
+
+
+ Problem-solving
+ Specialista di dominio + generalista
+
+
+
"Pensiamo passo dopo passo."
+
"Ragiona su questo con attenzione."
+
"Lavora su questo problema metodicamente."
+
"Mostra il tuo ragionamento ad ogni passo."
+
"Spiega il tuo processo di pensiero."
+
+
+
Usa CoT Per
+
+
**Problemi matematici** — Riduce errori di calcolo
+
**Puzzle logici** — Previene passi saltati
+
**Analisi complesse** — Organizza il pensiero
+
**Debug del codice** — Traccia l'esecuzione
+
**Processo decisionale** — Valuta trade-off
+
+
+
+
Salta CoT Per
+
+
**Q&A semplice** — Overhead non necessario
+
**Scrittura creativa** — Può limitare la creatività
+
**Ricerche fattuali** — Nessun ragionamento necessario
+
**Traduzione** — Compito diretto
+
**Riassunti** — Di solito semplici
+
+
+
+
+ Classificazione semplice
+ 2-3
+ Uno per categoria minimo
+
+
+ Formattazione complessa
+ 3-5
+ Mostra variazioni
+
+
+ Stile sfumato
+ 4-6
+ Cattura l'intera gamma
+
+
+ Casi limite
+ 1-2
+ Insieme agli esempi normali
+
+
+
+ Sintomo
+ Causa Probabile
+ Soluzione
+
+
+ Troppo lungo
+ Nessun vincolo di lunghezza
+ Aggiungi limiti parole/frasi
+
+
+ Troppo corto
+ Manca richiesta di dettagli
+ Chiedi elaborazione
+
+
+ Fuori tema
+ Istruzioni vaghe
+ Sii più specifico
+
+
+ Formato sbagliato
+ Formato non specificato
+ Definisci struttura esatta
+
+
+ Tono sbagliato
+ Pubblico non chiaro
+ Specifica pubblico/stile
+
+
+ Inconsistente
+ Nessun esempio fornito
+ Aggiungi esempi few-shot
+
+
+
+
Pronto per il Rilascio
+
+
L'output soddisfa consistentemente i requisiti
+
I casi limite sono gestiti appropriatamente
+
Il formato è affidabile e parsabile
+
Ulteriori miglioramenti mostrano rendimenti decrescenti
+
+
+
+
Continua a Iterare
+
+
L'output è inconsistente tra le esecuzioni
+
I casi limite causano fallimenti
+
Requisiti critici vengono mancati
+
Non hai testato abbastanza variazioni
+
+
+
+
+
Usa JSON Quando
+
+
Parsing programmatico necessario
+
Risposte API
+
Requisiti di tipo rigorosi
+
Integrazione JavaScript/Web
+
Rappresentazione compatta
+
+
+
+
Usa YAML Quando
+
+
La leggibilità umana conta
+
File di configurazione
+
Servono commenti
+
DevOps/Infrastruttura
+
Strutture profondamente annidate
+
+
+
+
+
Fare
+
+ - Dai un'identità chiara
+ - Elenca capacità specifiche
+ - Imposta confini espliciti
+ - Definisci il tono e lo stile
+ - Includi risposte di esempio
+
+
+
+
Non Fare
+
+ - Essere vago sul ruolo
+ - Dimenticare di impostare limiti
+ - Renderlo troppo lungo (max 500 parole)
+ - Contraddire te stesso
+ - Assumere che l'IA "capirà"
+
+
+
+
+
Difficoltà del Prompt Singolo
+
+
Il ragionamento multi-step si confonde
+
Diverse "modalità" di pensiero entrano in conflitto
+
Gli output complessi mancano di consistenza
+
Nessuna opportunità per controllo qualità
+
+
+
+
Il Chaining Risolve Questo
+
+
Ogni passo si concentra su un compito
+
Prompt specializzati per ogni modalità
+
Valida tra un passo e l'altro
+
Debug e migliora i singoli passi
+
+
+
+
+
→
+
+
+
Prompt 2
+
(Analizza)
+
+
Intermedio
+
+
→
+
+
+
Migliore Per
+
Riassunto documenti, generazione report, riutilizzo contenuti, conversione dati-a-narrativa
+
+
Migliore Per
+
Refactoring codice, pianificazione progetti, troubleshooting, decision-making strategico, problem solving complesso
+
+
Migliore Per
+
Copy marketing, scrittura creativa, bozze email, presentazioni, qualsiasi contenuto che beneficia di revisione
+
+
+
Reducing Latency
+
+
Parallelize independent steps
+
Cache intermediate results
+
Use smaller models for simple steps
+
Batch similar operations
+
+
+
+
Reducing Cost
+
+
Use cheaper models for classification
+
Limit iterations in loops
+
Short-circuit when possible
+
Cache repeated queries
+
+
+
+
Improving Reliability
+
+
Add validation between steps
+
Include retry logic
+
Log intermediate results
+
Implement fallback paths
+
+
+
+
+
Chaining Enables
+
+
Complex multi-step workflows
+
Higher quality through specialization
+
Better error handling and validation
+
Modular, reusable prompt components
+
+
+
+
Key Principles
+
+
Break complex tasks into simple steps
+
Design clear interfaces between steps
+
Validate intermediate outputs
+
Build in error handling and fallbacks
+
Optimize for your constraints
+
+
+
+
How RAG Works:
+
+
+ 1
+ User asks: "What's our refund policy?"
+
+
+ 2
+ System searches your documents for "refund policy"
+
+
+ 3
+ Finds relevant section from your policy document
+
+
+ 4
+ Sends to AI: "Based on this policy: [text], answer: What's our refund policy?"
+
+
+ 5
+ AI generates accurate answer using your actual policy
+
+
+
+
+
RAG Advantages
+
+ - Uses your actual, current data
+ - Reduces hallucinations
+ - Can cite sources
+ - Easy to update (just update documents)
+ - No expensive fine-tuning needed
+
+
+
+
When to Use RAG
+
+ - Customer support bots
+ - Documentation search
+ - Internal knowledge bases
+ - Any domain-specific Q&A
+ - When accuracy matters
+
+
+
+
+
System Prompts
+
Define the agent's identity, capabilities, constraints, and behavioral guidelines. These are the agent's "constitution."
+
+
+
Planning Prompts
+
Guide how agents break down complex goals into actionable steps. Critical for multi-step reasoning.
+
+
+
Tool-Use Prompts
+
Describe available tools and when/how to use them. Agents must understand their capabilities.
+
+
+
Reflection Prompts
+
Enable agents to evaluate their own outputs, catch errors, and improve iteratively.
+
+
+
+
+
→
+
+
→
+
+
→
+
(repeat)
+
+
+
+
Create Plan
+
Break goal into steps
+
+
↓
+
+
+
→
+
+
→
+
+
→
+
...
+
+
↓
+
+
Revise if Needed
+
Adapt plan based on results
+
+
+
+
+
+
Coordinator
+
Manages workflow
+
+
+
⟷
+
↕
+
+
+
+
+
↓
+
+
Planner Agent
+
Breaks down task
+
+
↓
+
+
Researcher Agent
+
Gathers information
+
+
↓
+
+
Writer Agent
+
Creates content
+
+
↓
+
+
Reviewer Agent
+
Quality checks
+
+
↓
+
+
+
+
+
+ | Today |
+ Tomorrow |
+
+
+
+
+ | Writing detailed prompts |
+ Designing agent systems |
+
+
+ | Manual prompt optimization |
+ Automated prompt tuning |
+
+
+ | Single-model expertise |
+ Multi-model orchestration |
+
+
+ | Text-focused interaction |
+ Multimodal fluency |
+
+
+ | Individual productivity |
+ Team-AI collaboration |
+
+
+
+
+
+
+
+
+
+
Agent
+
Autonomous AI system
+
+
+ powered by
+ ↓
+
+
+
+
+
+
+
Skill
+
Reusable expertise
+
+
+
+
+
+
Skill
+
Reusable expertise
+
+
+
+
+
+
Skill
+
Reusable expertise
+
+
+
+ composed of
+ ↓
+
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
Prompts are atoms → Skills are molecules → Agents are complete structures
+
+
+
+
+
+
→
+
+
+
+ ↺ Loop until complete
+
+
+
→
+
+
+
+
+
+
System Prompts
+
Identity & Role
+
+
+
Planning Prompts
+
How to Think
+
+
+
Tool Prompts
+
How to Act
+
+
+
Recovery Prompts
+
How to Recover
+
+
+
+
📄 SKILL.md (Required)
+
The main instruction file. Contains the core expertise, guidelines, and behaviors that define the skill.
+
+
+
📚 Reference Docs
+
Supporting documentation, examples, and context the agent can reference while working.
+
+
+
🔧 Scripts & Tools
+
Helper scripts, templates, or tool configurations that support the skill's functionality.
+
+
+
⚙️ Configuration
+
Settings, parameters, and customization options for adapting the skill to different contexts.
+
+
+
+ 📁
+ code-review-skill/
+
+
+
+ 📄
+ SKILL.md
+ Core review guidelines
+
+
+ 📄
+ security-checklist.md
+ Security patterns
+
+
+ 📄
+ performance-tips.md
+ Optimization guide
+
+
+
+ 📁
+ language-specific/
+
+
+
+ 📄
+ python.md
+ Python best practices
+
+
+ 📄
+ javascript.md
+ JavaScript patterns
+
+
+ 📄
+ rust.md
+ Rust guidelines
+
+
+
+
+
+
+
Simple Prompt
+
+
Single instruction
+
One-off use
+
Limited context
+
Generic approach
+
No supporting materials
+
+
+
+
Skill
+
+
Comprehensive instruction set
+
Reusable across projects
+
Rich context with references
+
Domain-specific expertise
+
Supporting docs, scripts, configs
+
+
+
+
+ Code Review
+
+
+
+
+ Security Audit
+
+
+
+
+ Documentation
+
+
=
+
+ Complete Code Quality Agent
+
+
+
+ AI Agent
+
+
+
+
Code Review
+
Skill 1
+
+
+
+
Test Writing
+
Skill 3
+
+
+
+
+
+
Core Prompts
+
Planning • Tools • Recovery • Memory
+
+
+
+
+
+
Fatih Kadir Akın
+
Creator of Awesome ChatGPT Prompts (a.k.a. prompts.chat)
+
+ Software developer from Istanbul, leading Developer Relations at Teknasyon. Author of books on JavaScript and prompt engineering. Open-source advocate specializing in web technologies and AI-assisted development.
+
+
+
+
+
+ GPT-4o
+ 128K tokens
+
+
+ GPT-5
+ 400K tokens
+
+
+ Claude Sonnet 4
+ 1M tokens
+
+
+ Gemini 2.5
+ 1M tokens
+
+
+ Llama 4
+ 1M-10M tokens
+
+
+ DeepSeek R1
+ 128K tokens
+
+
+
+ Quick questions
+ Faster models like GPT-4o or Claude 3.5 Sonnet
+
+
+ Hard problems
+ Smarter models like GPT-5.2 or Claude 4.5 Opus
+
+
+ Writing code
+ Code-focused models or the smartest general models
+
+
+ Long documents
+ Models with big context windows (Claude, Gemini)
+
+
+ Current events
+ Models with internet access
+
+
+
+ Creation
+ Write, Create, Generate, Compose, Design
+
+
+ Analysis
+ Analyze, Evaluate, Compare, Assess, Review
+
+
+ Transformation
+ Convert, Translate, Reformat, Summarize, Expand
+
+
+ Explanation
+ Explain, Describe, Clarify, Define, Illustrate
+
+
+ Problem-solving
+ Solve, Debug, Fix, Optimize, Improve
+
+
+
+ Audience
+ Who will read/use this?
+
+
+ Length
+ How long/short should it be?
+
+
+ Tone
+ Formal? Casual? Technical?
+
+
+ Format
+ Prose? List? Table? Code?
+
+
+ Scope
+ What to include/exclude?
+
+
+ Purpose
+ What should this accomplish?
+
+
+
+ Too verbose
+ Add "Be concise" or length limits
+
+
+ Too vague
+ Add specific examples or constraints
+
+
+ Wrong format
+ Specify exact output structure
+
+
+ Missing aspects
+ Add "Make sure to include..."
+
+
+ Wrong tone
+ Specify audience and style
+
+
+ Inaccurate
+ Request citations or step-by-step reasoning
+
+
+
+ Code review
+ Senior developer + mentor
+
+
+ Writing feedback
+ Editor + target audience member
+
+
+ Business strategy
+ Consultant + industry expert
+
+
+ Learning new topic
+ Patient teacher + practitioner
+
+
+ Creative writing
+ Specific genre author
+
+
+ Technical explanation
+ Expert + communicator
+
+
+ Problem-solving
+ Domain specialist + generalist
+
+
+
"Let's think step by step."
+
"Think through this carefully."
+
"Work through this problem methodically."
+
"Show your reasoning at each step."
+
"Explain your thought process."
+
+
+
Use CoT For
+
+
**Math problems** — Reduces calculation errors
+
**Logic puzzles** — Prevents skipped steps
+
**Complex analysis** — Organizes thinking
+
**Code debugging** — Traces execution
+
**Decision making** — Weighs trade-offs
+
+
+
+
Skip CoT For
+
+
**Simple Q&A** — Unnecessary overhead
+
**Creative writing** — Can constrain creativity
+
**Factual lookups** — No reasoning needed
+
**Translation** — Direct task
+
**Summarization** — Usually straightforward
+
+
+
+
+ Simple classification
+ 2-3
+ One per category minimum
+
+
+ Complex formatting
+ 3-5
+ Show variations
+
+
+ Nuanced style
+ 4-6
+ Capture full range
+
+
+ Edge cases
+ 1-2
+ Alongside normal examples
+
+
+
+ Symptom
+ Likely Cause
+ Solution
+
+
+ Too long
+ No length constraint
+ Add word/sentence limits
+
+
+ Too short
+ Lacks detail request
+ Ask for elaboration
+
+
+ Off-topic
+ Vague instructions
+ Be more specific
+
+
+ Wrong format
+ Format not specified
+ Define exact structure
+
+
+ Wrong tone
+ Audience not clear
+ Specify audience/style
+
+
+ Inconsistent
+ No examples provided
+ Add few-shot examples
+
+
+
+
Ready to Ship
+
+
Output consistently meets requirements
+
Edge cases are handled appropriately
+
Format is reliable and parseable
+
Further improvements show diminishing returns
+
+
+
+
Keep Iterating
+
+
Output is inconsistent across runs
+
Edge cases cause failures
+
Critical requirements are missed
+
You haven't tested enough variations
+
+
+
+
+
Use JSON When
+
+
Programmatic parsing needed
+
API responses
+
Strict type requirements
+
JavaScript/Web integration
+
Compact representation
+
+
+
+
Use YAML When
+
+
Human readability matters
+
Configuration files
+
Comments are needed
+
DevOps/Infrastructure
+
Deep nested structures
+
+
+
+
+
Do
+
+ - Give a clear identity
+ - List specific capabilities
+ - Set explicit boundaries
+ - Define the tone and style
+ - Include example responses
+
+
+
+
Don't
+
+ - Be vague about the role
+ - Forget to set limits
+ - Make it too long (500 words max)
+ - Contradict yourself
+ - Assume the AI will "figure it out"
+
+
+
+
+
Single Prompt Struggles
+
+
Multi-step reasoning gets confused
+
Different "modes" of thinking clash
+
Complex outputs lack consistency
+
No opportunity for quality control
+
+
+
+
Chaining Solves This
+
+
Each step focuses on one task
+
Specialized prompts for each mode
+
Validate between steps
+
Debug and improve individual steps
+
+
+
+
+
→
+
+
→
+
+
+
Prompt 3
+
(Generate)
+
+
Output
+
+
+
Best For
+
Document summarization, report generation, content repurposing, data-to-narrative conversion
+
+
Best For
+
Code refactoring, project planning, troubleshooting, strategic decision-making, complex problem solving
+
+
Best For
+
Marketing copy, creative writing, email drafts, presentations, any content that benefits from revision
+
+
+
Reducing Latency
+
+
Parallelize independent steps
+
Cache intermediate results
+
Use smaller models for simple steps
+
Batch similar operations
+
+
+
+
Reducing Cost
+
+
Use cheaper models for classification
+
Limit iterations in loops
+
Short-circuit when possible
+
Cache repeated queries
+
+
+
+
Improving Reliability
+
+
Add validation between steps
+
Include retry logic
+
Log intermediate results
+
Implement fallback paths
+
+
+
+
+
Chaining Enables
+
+
Complex multi-step workflows
+
Higher quality through specialization
+
Better error handling and validation
+
Modular, reusable prompt components
+
+
+
+
Key Principles
+
+
Break complex tasks into simple steps
+
Design clear interfaces between steps
+
Validate intermediate outputs
+
Build in error handling and fallbacks
+
Optimize for your constraints
+
+
+
+
How RAG Works:
+
+
+ 1
+ User asks: "What's our refund policy?"
+
+
+ 2
+ System searches your documents for "refund policy"
+
+
+ 3
+ Finds relevant section from your policy document
+
+
+ 4
+ Sends to AI: "Based on this policy: [text], answer: What's our refund policy?"
+
+
+ 5
+ AI generates accurate answer using your actual policy
+
+
+
+
+
RAG Advantages
+
+ - Uses your actual, current data
+ - Reduces hallucinations
+ - Can cite sources
+ - Easy to update (just update documents)
+ - No expensive fine-tuning needed
+
+
+
+
When to Use RAG
+
+ - Customer support bots
+ - Documentation search
+ - Internal knowledge bases
+ - Any domain-specific Q&A
+ - When accuracy matters
+
+
+
+
+
System Prompts
+
Define the agent's identity, capabilities, constraints, and behavioral guidelines. These are the agent's "constitution."
+
+
+
Planning Prompts
+
Guide how agents break down complex goals into actionable steps. Critical for multi-step reasoning.
+
+
+
Tool-Use Prompts
+
Describe available tools and when/how to use them. Agents must understand their capabilities.
+
+
+
Reflection Prompts
+
Enable agents to evaluate their own outputs, catch errors, and improve iteratively.
+
+
+
+
+
→
+
+
→
+
+
→
+
(repeat)
+
+
+
+
Create Plan
+
Break goal into steps
+
+
↓
+
+
+
→
+
+
→
+
+
→
+
...
+
+
↓
+
+
Revise if Needed
+
Adapt plan based on results
+
+
+
+
+
+
Coordinator
+
Manages workflow
+
+
+
⟷
+
↕
+
+
+
+
+
↓
+
+
Planner Agent
+
Breaks down task
+
+
↓
+
+
Researcher Agent
+
Gathers information
+
+
↓
+
+
Writer Agent
+
Creates content
+
+
↓
+
+
Reviewer Agent
+
Quality checks
+
+
↓
+
+
+
+
+
+ | Today |
+ Tomorrow |
+
+
+
+
+ | Writing detailed prompts |
+ Designing agent systems |
+
+
+ | Manual prompt optimization |
+ Automated prompt tuning |
+
+
+ | Single-model expertise |
+ Multi-model orchestration |
+
+
+ | Text-focused interaction |
+ Multimodal fluency |
+
+
+ | Individual productivity |
+ Team-AI collaboration |
+
+
+
+
+
+
+
+
+
+
Agent
+
Autonomous AI system
+
+
+ powered by
+ ↓
+
+
+
+
+
+
+
Skill
+
Reusable expertise
+
+
+
+
+
+
Skill
+
Reusable expertise
+
+
+
+
+
+
Skill
+
Reusable expertise
+
+
+
+ composed of
+ ↓
+
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
Prompts are atoms → Skills are molecules → Agents are complete structures
+
+
+
+
+
+
→
+
+
+
+ ↺ Loop until complete
+
+
+
→
+
+
+
+
+
+
System Prompts
+
Identity & Role
+
+
+
Planning Prompts
+
How to Think
+
+
+
Tool Prompts
+
How to Act
+
+
+
Recovery Prompts
+
How to Recover
+
+
+
+
📄 SKILL.md (Required)
+
The main instruction file. Contains the core expertise, guidelines, and behaviors that define the skill.
+
+
+
📚 Reference Docs
+
Supporting documentation, examples, and context the agent can reference while working.
+
+
+
🔧 Scripts & Tools
+
Helper scripts, templates, or tool configurations that support the skill's functionality.
+
+
+
⚙️ Configuration
+
Settings, parameters, and customization options for adapting the skill to different contexts.
+
+
+
+ 📁
+ code-review-skill/
+
+
+
+ 📄
+ SKILL.md
+ Core review guidelines
+
+
+ 📄
+ security-checklist.md
+ Security patterns
+
+
+ 📄
+ performance-tips.md
+ Optimization guide
+
+
+
+ 📁
+ language-specific/
+
+
+
+ 📄
+ python.md
+ Python best practices
+
+
+ 📄
+ javascript.md
+ JavaScript patterns
+
+
+ 📄
+ rust.md
+ Rust guidelines
+
+
+
+
+
+
+
Simple Prompt
+
+
Single instruction
+
One-off use
+
Limited context
+
Generic approach
+
No supporting materials
+
+
+
+
Skill
+
+
Comprehensive instruction set
+
Reusable across projects
+
Rich context with references
+
Domain-specific expertise
+
Supporting docs, scripts, configs
+
+
+
+
+ Code Review
+
+
+
+
+ Security Audit
+
+
+
+
+ Documentation
+
+
=
+
+ Complete Code Quality Agent
+
+
+
+ AI Agent
+
+
+
+
Code Review
+
Skill 1
+
+
+
+
Test Writing
+
Skill 3
+
+
+
+
+
+
Core Prompts
+
Planning • Tools • Recovery • Memory
+
+
+
+
+
+
Fatih Kadir Akın
+
Creator of Awesome ChatGPT Prompts (a.k.a. prompts.chat)
+
+ Software developer from Istanbul, leading Developer Relations at Teknasyon. Author of books on JavaScript and prompt engineering. Open-source advocate specializing in web technologies and AI-assisted development.
+
+
+
+
+
+ GPT-4o
+ 128K tokens
+
+
+ GPT-5
+ 400K tokens
+
+
+ Claude Sonnet 4
+ 1M tokens
+
+
+ Gemini 2.5
+ 1M tokens
+
+
+ Llama 4
+ 1M-10M tokens
+
+
+ DeepSeek R1
+ 128K tokens
+
+
+
+ Quick questions
+ Faster models like GPT-4o or Claude 3.5 Sonnet
+
+
+ Hard problems
+ Smarter models like GPT-5.2 or Claude 4.5 Opus
+
+
+ Writing code
+ Code-focused models or the smartest general models
+
+
+ Long documents
+ Models with big context windows (Claude, Gemini)
+
+
+ Current events
+ Models with internet access
+
+
+
+ Creation
+ Write, Create, Generate, Compose, Design
+
+
+ Analysis
+ Analyze, Evaluate, Compare, Assess, Review
+
+
+ Transformation
+ Convert, Translate, Reformat, Summarize, Expand
+
+
+ Explanation
+ Explain, Describe, Clarify, Define, Illustrate
+
+
+ Problem-solving
+ Solve, Debug, Fix, Optimize, Improve
+
+
+
+ Audience
+ Who will read/use this?
+
+
+ Length
+ How long/short should it be?
+
+
+ Tone
+ Formal? Casual? Technical?
+
+
+ Format
+ Prose? List? Table? Code?
+
+
+ Scope
+ What to include/exclude?
+
+
+ Purpose
+ What should this accomplish?
+
+
+
+ Too verbose
+ Add "Be concise" or length limits
+
+
+ Too vague
+ Add specific examples or constraints
+
+
+ Wrong format
+ Specify exact output structure
+
+
+ Missing aspects
+ Add "Make sure to include..."
+
+
+ Wrong tone
+ Specify audience and style
+
+
+ Inaccurate
+ Request citations or step-by-step reasoning
+
+
+
+ Code review
+ Senior developer + mentor
+
+
+ Writing feedback
+ Editor + target audience member
+
+
+ Business strategy
+ Consultant + industry expert
+
+
+ Learning new topic
+ Patient teacher + practitioner
+
+
+ Creative writing
+ Specific genre author
+
+
+ Technical explanation
+ Expert + communicator
+
+
+ Problem-solving
+ Domain specialist + generalist
+
+
+
"Let's think step by step."
+
"Think through this carefully."
+
"Work through this problem methodically."
+
"Show your reasoning at each step."
+
"Explain your thought process."
+
+
+
Use CoT For
+
+
**Math problems** — Reduces calculation errors
+
**Logic puzzles** — Prevents skipped steps
+
**Complex analysis** — Organizes thinking
+
**Code debugging** — Traces execution
+
**Decision making** — Weighs trade-offs
+
+
+
+
Skip CoT For
+
+
**Simple Q&A** — Unnecessary overhead
+
**Creative writing** — Can constrain creativity
+
**Factual lookups** — No reasoning needed
+
**Translation** — Direct task
+
**Summarization** — Usually straightforward
+
+
+
+
+ Simple classification
+ 2-3
+ One per category minimum
+
+
+ Complex formatting
+ 3-5
+ Show variations
+
+
+ Nuanced style
+ 4-6
+ Capture full range
+
+
+ Edge cases
+ 1-2
+ Alongside normal examples
+
+
+
+ Symptom
+ Likely Cause
+ Solution
+
+
+ Too long
+ No length constraint
+ Add word/sentence limits
+
+
+ Too short
+ Lacks detail request
+ Ask for elaboration
+
+
+ Off-topic
+ Vague instructions
+ Be more specific
+
+
+ Wrong format
+ Format not specified
+ Define exact structure
+
+
+ Wrong tone
+ Audience not clear
+ Specify audience/style
+
+
+ Inconsistent
+ No examples provided
+ Add few-shot examples
+
+
+
+
Ready to Ship
+
+
Output consistently meets requirements
+
Edge cases are handled appropriately
+
Format is reliable and parseable
+
Further improvements show diminishing returns
+
+
+
+
Keep Iterating
+
+
Output is inconsistent across runs
+
Edge cases cause failures
+
Critical requirements are missed
+
You haven't tested enough variations
+
+
+
+
+
Use JSON When
+
+
Programmatic parsing needed
+
API responses
+
Strict type requirements
+
JavaScript/Web integration
+
Compact representation
+
+
+
+
Use YAML When
+
+
Human readability matters
+
Configuration files
+
Comments are needed
+
DevOps/Infrastructure
+
Deep nested structures
+
+
+
+
+
Do
+
+ - Give a clear identity
+ - List specific capabilities
+ - Set explicit boundaries
+ - Define the tone and style
+ - Include example responses
+
+
+
+
Don't
+
+ - Be vague about the role
+ - Forget to set limits
+ - Make it too long (500 words max)
+ - Contradict yourself
+ - Assume the AI will "figure it out"
+
+
+
+
+
Single Prompt Struggles
+
+
Multi-step reasoning gets confused
+
Different "modes" of thinking clash
+
Complex outputs lack consistency
+
No opportunity for quality control
+
+
+
+
Chaining Solves This
+
+
Each step focuses on one task
+
Specialized prompts for each mode
+
Validate between steps
+
Debug and improve individual steps
+
+
+
+
+
→
+
+
→
+
+
+
Prompt 3
+
(Generate)
+
+
Output
+
+
+
Best For
+
Document summarization, report generation, content repurposing, data-to-narrative conversion
+
+
Best For
+
Code refactoring, project planning, troubleshooting, strategic decision-making, complex problem solving
+
+
Best For
+
Marketing copy, creative writing, email drafts, presentations, any content that benefits from revision
+
+
+
Reducing Latency
+
+
Parallelize independent steps
+
Cache intermediate results
+
Use smaller models for simple steps
+
Batch similar operations
+
+
+
+
Reducing Cost
+
+
Use cheaper models for classification
+
Limit iterations in loops
+
Short-circuit when possible
+
Cache repeated queries
+
+
+
+
Improving Reliability
+
+
Add validation between steps
+
Include retry logic
+
Log intermediate results
+
Implement fallback paths
+
+
+
+
+
Chaining Enables
+
+
Complex multi-step workflows
+
Higher quality through specialization
+
Better error handling and validation
+
Modular, reusable prompt components
+
+
+
+
Key Principles
+
+
Break complex tasks into simple steps
+
Design clear interfaces between steps
+
Validate intermediate outputs
+
Build in error handling and fallbacks
+
Optimize for your constraints
+
+
+
+
How RAG Works:
+
+
+ 1
+ User asks: "What's our refund policy?"
+
+
+ 2
+ System searches your documents for "refund policy"
+
+
+ 3
+ Finds relevant section from your policy document
+
+
+ 4
+ Sends to AI: "Based on this policy: [text], answer: What's our refund policy?"
+
+
+ 5
+ AI generates accurate answer using your actual policy
+
+
+
+
+
RAG Advantages
+
+ - Uses your actual, current data
+ - Reduces hallucinations
+ - Can cite sources
+ - Easy to update (just update documents)
+ - No expensive fine-tuning needed
+
+
+
+
When to Use RAG
+
+ - Customer support bots
+ - Documentation search
+ - Internal knowledge bases
+ - Any domain-specific Q&A
+ - When accuracy matters
+
+
+
+
+
System Prompts
+
Define the agent's identity, capabilities, constraints, and behavioral guidelines. These are the agent's "constitution."
+
+
+
Planning Prompts
+
Guide how agents break down complex goals into actionable steps. Critical for multi-step reasoning.
+
+
+
Tool-Use Prompts
+
Describe available tools and when/how to use them. Agents must understand their capabilities.
+
+
+
Reflection Prompts
+
Enable agents to evaluate their own outputs, catch errors, and improve iteratively.
+
+
+
+
+
→
+
+
→
+
+
→
+
(repeat)
+
+
+
+
Create Plan
+
Break goal into steps
+
+
↓
+
+
+
→
+
+
→
+
+
→
+
...
+
+
↓
+
+
Revise if Needed
+
Adapt plan based on results
+
+
+
+
+
+
Coordinator
+
Manages workflow
+
+
+
⟷
+
↕
+
+
+
+
+
↓
+
+
Planner Agent
+
Breaks down task
+
+
↓
+
+
Researcher Agent
+
Gathers information
+
+
↓
+
+
Writer Agent
+
Creates content
+
+
↓
+
+
Reviewer Agent
+
Quality checks
+
+
↓
+
+
+
+
+
+ | Today |
+ Tomorrow |
+
+
+
+
+ | Writing detailed prompts |
+ Designing agent systems |
+
+
+ | Manual prompt optimization |
+ Automated prompt tuning |
+
+
+ | Single-model expertise |
+ Multi-model orchestration |
+
+
+ | Text-focused interaction |
+ Multimodal fluency |
+
+
+ | Individual productivity |
+ Team-AI collaboration |
+
+
+
+
+
+
+
+
+
+
Agent
+
Autonomous AI system
+
+
+ powered by
+ ↓
+
+
+
+
+
+
+
Skill
+
Reusable expertise
+
+
+
+
+
+
Skill
+
Reusable expertise
+
+
+
+
+
+
Skill
+
Reusable expertise
+
+
+
+ composed of
+ ↓
+
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
Prompts are atoms → Skills are molecules → Agents are complete structures
+
+
+
+
+
+
→
+
+
+
+ ↺ Loop until complete
+
+
+
→
+
+
+
+
+
+
System Prompts
+
Identity & Role
+
+
+
Planning Prompts
+
How to Think
+
+
+
Tool Prompts
+
How to Act
+
+
+
Recovery Prompts
+
How to Recover
+
+
+
+
📄 SKILL.md (Required)
+
The main instruction file. Contains the core expertise, guidelines, and behaviors that define the skill.
+
+
+
📚 Reference Docs
+
Supporting documentation, examples, and context the agent can reference while working.
+
+
+
🔧 Scripts & Tools
+
Helper scripts, templates, or tool configurations that support the skill's functionality.
+
+
+
⚙️ Configuration
+
Settings, parameters, and customization options for adapting the skill to different contexts.
+
+
+
+ 📁
+ code-review-skill/
+
+
+
+ 📄
+ SKILL.md
+ Core review guidelines
+
+
+ 📄
+ security-checklist.md
+ Security patterns
+
+
+ 📄
+ performance-tips.md
+ Optimization guide
+
+
+
+ 📁
+ language-specific/
+
+
+
+ 📄
+ python.md
+ Python best practices
+
+
+ 📄
+ javascript.md
+ JavaScript patterns
+
+
+ 📄
+ rust.md
+ Rust guidelines
+
+
+
+
+
+
+
Simple Prompt
+
+
Single instruction
+
One-off use
+
Limited context
+
Generic approach
+
No supporting materials
+
+
+
+
Skill
+
+
Comprehensive instruction set
+
Reusable across projects
+
Rich context with references
+
Domain-specific expertise
+
Supporting docs, scripts, configs
+
+
+
+
+ Code Review
+
+
+
+
+ Security Audit
+
+
+
+
+ Documentation
+
+
=
+
+ Complete Code Quality Agent
+
+
+
+ AI Agent
+
+
+
+
Code Review
+
Skill 1
+
+
+
+
Test Writing
+
Skill 3
+
+
+
+
+
+
Core Prompts
+
Planning • Tools • Recovery • Memory
+
+
+
+
+
+
Fatih Kadir Akın
+
Creator of Awesome ChatGPT Prompts (a.k.a. prompts.chat)
+
+ Software developer from Istanbul, leading Developer Relations at Teknasyon. Author of books on JavaScript and prompt engineering. Open-source advocate specializing in web technologies and AI-assisted development.
+
+
+
+
+
+ GPT-4o
+ 128K tokens
+
+
+ GPT-5
+ 400K tokens
+
+
+ Claude Sonnet 4
+ 1M tokens
+
+
+ Gemini 2.5
+ 1M tokens
+
+
+ Llama 4
+ 1M-10M tokens
+
+
+ DeepSeek R1
+ 128K tokens
+
+
+
+ Quick questions
+ Faster models like GPT-4o or Claude 3.5 Sonnet
+
+
+ Hard problems
+ Smarter models like GPT-5.2 or Claude 4.5 Opus
+
+
+ Writing code
+ Code-focused models or the smartest general models
+
+
+ Long documents
+ Models with big context windows (Claude, Gemini)
+
+
+ Current events
+ Models with internet access
+
+
+
+ Creation
+ Write, Create, Generate, Compose, Design
+
+
+ Analysis
+ Analyze, Evaluate, Compare, Assess, Review
+
+
+ Transformation
+ Convert, Translate, Reformat, Summarize, Expand
+
+
+ Explanation
+ Explain, Describe, Clarify, Define, Illustrate
+
+
+ Problem-solving
+ Solve, Debug, Fix, Optimize, Improve
+
+
+
+ Audience
+ Who will read/use this?
+
+
+ Length
+ How long/short should it be?
+
+
+ Tone
+ Formal? Casual? Technical?
+
+
+ Format
+ Prose? List? Table? Code?
+
+
+ Scope
+ What to include/exclude?
+
+
+ Purpose
+ What should this accomplish?
+
+
+
+ Too verbose
+ Add "Be concise" or length limits
+
+
+ Too vague
+ Add specific examples or constraints
+
+
+ Wrong format
+ Specify exact output structure
+
+
+ Missing aspects
+ Add "Make sure to include..."
+
+
+ Wrong tone
+ Specify audience and style
+
+
+ Inaccurate
+ Request citations or step-by-step reasoning
+
+
+
+ Code review
+ Senior developer + mentor
+
+
+ Writing feedback
+ Editor + target audience member
+
+
+ Business strategy
+ Consultant + industry expert
+
+
+ Learning new topic
+ Patient teacher + practitioner
+
+
+ Creative writing
+ Specific genre author
+
+
+ Technical explanation
+ Expert + communicator
+
+
+ Problem-solving
+ Domain specialist + generalist
+
+
+
"Let's think step by step."
+
"Think through this carefully."
+
"Work through this problem methodically."
+
"Show your reasoning at each step."
+
"Explain your thought process."
+
+
+
Use CoT For
+
+
**Math problems** — Reduces calculation errors
+
**Logic puzzles** — Prevents skipped steps
+
**Complex analysis** — Organizes thinking
+
**Code debugging** — Traces execution
+
**Decision making** — Weighs trade-offs
+
+
+
+
Skip CoT For
+
+
**Simple Q&A** — Unnecessary overhead
+
**Creative writing** — Can constrain creativity
+
**Factual lookups** — No reasoning needed
+
**Translation** — Direct task
+
**Summarization** — Usually straightforward
+
+
+
+
+ Simple classification
+ 2-3
+ One per category minimum
+
+
+ Complex formatting
+ 3-5
+ Show variations
+
+
+ Nuanced style
+ 4-6
+ Capture full range
+
+
+ Edge cases
+ 1-2
+ Alongside normal examples
+
+
+
+ Symptom
+ Likely Cause
+ Solution
+
+
+ Too long
+ No length constraint
+ Add word/sentence limits
+
+
+ Too short
+ Lacks detail request
+ Ask for elaboration
+
+
+ Off-topic
+ Vague instructions
+ Be more specific
+
+
+ Wrong format
+ Format not specified
+ Define exact structure
+
+
+ Wrong tone
+ Audience not clear
+ Specify audience/style
+
+
+ Inconsistent
+ No examples provided
+ Add few-shot examples
+
+
+
+
Ready to Ship
+
+
Output consistently meets requirements
+
Edge cases are handled appropriately
+
Format is reliable and parseable
+
Further improvements show diminishing returns
+
+
+
+
Keep Iterating
+
+
Output is inconsistent across runs
+
Edge cases cause failures
+
Critical requirements are missed
+
You haven't tested enough variations
+
+
+
+
+
Use JSON When
+
+
Programmatic parsing needed
+
API responses
+
Strict type requirements
+
JavaScript/Web integration
+
Compact representation
+
+
+
+
Use YAML When
+
+
Human readability matters
+
Configuration files
+
Comments are needed
+
DevOps/Infrastructure
+
Deep nested structures
+
+
+
+
+
Do
+
+ - Give a clear identity
+ - List specific capabilities
+ - Set explicit boundaries
+ - Define the tone and style
+ - Include example responses
+
+
+
+
Don't
+
+ - Be vague about the role
+ - Forget to set limits
+ - Make it too long (500 words max)
+ - Contradict yourself
+ - Assume the AI will "figure it out"
+
+
+
+
+
Single Prompt Struggles
+
+
Multi-step reasoning gets confused
+
Different "modes" of thinking clash
+
Complex outputs lack consistency
+
No opportunity for quality control
+
+
+
+
Chaining Solves This
+
+
Each step focuses on one task
+
Specialized prompts for each mode
+
Validate between steps
+
Debug and improve individual steps
+
+
+
+
+
→
+
+
→
+
+
+
Prompt 3
+
(Generate)
+
+
Output
+
+
+
Best For
+
Document summarization, report generation, content repurposing, data-to-narrative conversion
+
+
Best For
+
Code refactoring, project planning, troubleshooting, strategic decision-making, complex problem solving
+
+
Best For
+
Marketing copy, creative writing, email drafts, presentations, any content that benefits from revision
+
+
+
Reducing Latency
+
+
Parallelize independent steps
+
Cache intermediate results
+
Use smaller models for simple steps
+
Batch similar operations
+
+
+
+
Reducing Cost
+
+
Use cheaper models for classification
+
Limit iterations in loops
+
Short-circuit when possible
+
Cache repeated queries
+
+
+
+
Improving Reliability
+
+
Add validation between steps
+
Include retry logic
+
Log intermediate results
+
Implement fallback paths
+
+
+
+
+
Chaining Enables
+
+
Complex multi-step workflows
+
Higher quality through specialization
+
Better error handling and validation
+
Modular, reusable prompt components
+
+
+
+
Key Principles
+
+
Break complex tasks into simple steps
+
Design clear interfaces between steps
+
Validate intermediate outputs
+
Build in error handling and fallbacks
+
Optimize for your constraints
+
+
+
+
How RAG Works:
+
+
+ 1
+ User asks: "What's our refund policy?"
+
+
+ 2
+ System searches your documents for "refund policy"
+
+
+ 3
+ Finds relevant section from your policy document
+
+
+ 4
+ Sends to AI: "Based on this policy: [text], answer: What's our refund policy?"
+
+
+ 5
+ AI generates accurate answer using your actual policy
+
+
+
+
+
RAG Advantages
+
+ - Uses your actual, current data
+ - Reduces hallucinations
+ - Can cite sources
+ - Easy to update (just update documents)
+ - No expensive fine-tuning needed
+
+
+
+
When to Use RAG
+
+ - Customer support bots
+ - Documentation search
+ - Internal knowledge bases
+ - Any domain-specific Q&A
+ - When accuracy matters
+
+
+
+
+
System Prompts
+
Define the agent's identity, capabilities, constraints, and behavioral guidelines. These are the agent's "constitution."
+
+
+
Planning Prompts
+
Guide how agents break down complex goals into actionable steps. Critical for multi-step reasoning.
+
+
+
Tool-Use Prompts
+
Describe available tools and when/how to use them. Agents must understand their capabilities.
+
+
+
Reflection Prompts
+
Enable agents to evaluate their own outputs, catch errors, and improve iteratively.
+
+
+
+
+
→
+
+
→
+
+
→
+
(repeat)
+
+
+
+
Create Plan
+
Break goal into steps
+
+
↓
+
+
+
→
+
+
→
+
+
→
+
...
+
+
↓
+
+
Revise if Needed
+
Adapt plan based on results
+
+
+
+
+
+
Coordinator
+
Manages workflow
+
+
+
⟷
+
↕
+
+
+
+
+
↓
+
+
Planner Agent
+
Breaks down task
+
+
↓
+
+
Researcher Agent
+
Gathers information
+
+
↓
+
+
Writer Agent
+
Creates content
+
+
↓
+
+
Reviewer Agent
+
Quality checks
+
+
↓
+
+
+
+
+
+ | Today |
+ Tomorrow |
+
+
+
+
+ | Writing detailed prompts |
+ Designing agent systems |
+
+
+ | Manual prompt optimization |
+ Automated prompt tuning |
+
+
+ | Single-model expertise |
+ Multi-model orchestration |
+
+
+ | Text-focused interaction |
+ Multimodal fluency |
+
+
+ | Individual productivity |
+ Team-AI collaboration |
+
+
+
+
+
+
+
+
+
+
Agent
+
Autonomous AI system
+
+
+ powered by
+ ↓
+
+
+
+
+
+
+
Skill
+
Reusable expertise
+
+
+
+
+
+
Skill
+
Reusable expertise
+
+
+
+
+
+
Skill
+
Reusable expertise
+
+
+
+ composed of
+ ↓
+
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
Prompts are atoms → Skills are molecules → Agents are complete structures
+
+
+
+
+
+
→
+
+
+
+ ↺ Loop until complete
+
+
+
→
+
+
+
+
+
+
System Prompts
+
Identity & Role
+
+
+
Planning Prompts
+
How to Think
+
+
+
Tool Prompts
+
How to Act
+
+
+
Recovery Prompts
+
How to Recover
+
+
+
+
📄 SKILL.md (Required)
+
The main instruction file. Contains the core expertise, guidelines, and behaviors that define the skill.
+
+
+
📚 Reference Docs
+
Supporting documentation, examples, and context the agent can reference while working.
+
+
+
🔧 Scripts & Tools
+
Helper scripts, templates, or tool configurations that support the skill's functionality.
+
+
+
⚙️ Configuration
+
Settings, parameters, and customization options for adapting the skill to different contexts.
+
+
+
+ 📁
+ code-review-skill/
+
+
+
+ 📄
+ SKILL.md
+ Core review guidelines
+
+
+ 📄
+ security-checklist.md
+ Security patterns
+
+
+ 📄
+ performance-tips.md
+ Optimization guide
+
+
+
+ 📁
+ language-specific/
+
+
+
+ 📄
+ python.md
+ Python best practices
+
+
+ 📄
+ javascript.md
+ JavaScript patterns
+
+
+ 📄
+ rust.md
+ Rust guidelines
+
+
+
+
+
+
+
Simple Prompt
+
+
Single instruction
+
One-off use
+
Limited context
+
Generic approach
+
No supporting materials
+
+
+
+
Skill
+
+
Comprehensive instruction set
+
Reusable across projects
+
Rich context with references
+
Domain-specific expertise
+
Supporting docs, scripts, configs
+
+
+
+
+ Code Review
+
+
+
+
+ Security Audit
+
+
+
+
+ Documentation
+
+
=
+
+ Complete Code Quality Agent
+
+
+
+ AI Agent
+
+
+
+
Code Review
+
Skill 1
+
+
+
+
Test Writing
+
Skill 3
+
+
+
+
+
+
Core Prompts
+
Planning • Tools • Recovery • Memory
+
+
+
+
+
+
Fatih Kadir Akın
+
Creator of Awesome ChatGPT Prompts (a.k.a. prompts.chat)
+
+ Software developer from Istanbul, leading Developer Relations at Teknasyon. Author of books on JavaScript and prompt engineering. Open-source advocate specializing in web technologies and AI-assisted development.
+
+
+
+
+
+ GPT-4o
+ 128K tokens
+
+
+ GPT-5
+ 400K tokens
+
+
+ Claude Sonnet 4
+ 1M tokens
+
+
+ Gemini 2.5
+ 1M tokens
+
+
+ Llama 4
+ 1M-10M tokens
+
+
+ DeepSeek R1
+ 128K tokens
+
+
+
+ Quick questions
+ Faster models like GPT-4o or Claude 3.5 Sonnet
+
+
+ Hard problems
+ Smarter models like GPT-5.2 or Claude 4.5 Opus
+
+
+ Writing code
+ Code-focused models or the smartest general models
+
+
+ Long documents
+ Models with big context windows (Claude, Gemini)
+
+
+ Current events
+ Models with internet access
+
+
+
+ Creation
+ Write, Create, Generate, Compose, Design
+
+
+ Analysis
+ Analyze, Evaluate, Compare, Assess, Review
+
+
+ Transformation
+ Convert, Translate, Reformat, Summarize, Expand
+
+
+ Explanation
+ Explain, Describe, Clarify, Define, Illustrate
+
+
+ Problem-solving
+ Solve, Debug, Fix, Optimize, Improve
+
+
+
+ Audience
+ Who will read/use this?
+
+
+ Length
+ How long/short should it be?
+
+
+ Tone
+ Formal? Casual? Technical?
+
+
+ Format
+ Prose? List? Table? Code?
+
+
+ Scope
+ What to include/exclude?
+
+
+ Purpose
+ What should this accomplish?
+
+
+
+ Too verbose
+ Add "Be concise" or length limits
+
+
+ Too vague
+ Add specific examples or constraints
+
+
+ Wrong format
+ Specify exact output structure
+
+
+ Missing aspects
+ Add "Make sure to include..."
+
+
+ Wrong tone
+ Specify audience and style
+
+
+ Inaccurate
+ Request citations or step-by-step reasoning
+
+
+
+ Code review
+ Senior developer + mentor
+
+
+ Writing feedback
+ Editor + target audience member
+
+
+ Business strategy
+ Consultant + industry expert
+
+
+ Learning new topic
+ Patient teacher + practitioner
+
+
+ Creative writing
+ Specific genre author
+
+
+ Technical explanation
+ Expert + communicator
+
+
+ Problem-solving
+ Domain specialist + generalist
+
+
+
"Let's think step by step."
+
"Think through this carefully."
+
"Work through this problem methodically."
+
"Show your reasoning at each step."
+
"Explain your thought process."
+
+
+
Use CoT For
+
+
**Math problems** — Reduces calculation errors
+
**Logic puzzles** — Prevents skipped steps
+
**Complex analysis** — Organizes thinking
+
**Code debugging** — Traces execution
+
**Decision making** — Weighs trade-offs
+
+
+
+
Skip CoT For
+
+
**Simple Q&A** — Unnecessary overhead
+
**Creative writing** — Can constrain creativity
+
**Factual lookups** — No reasoning needed
+
**Translation** — Direct task
+
**Summarization** — Usually straightforward
+
+
+
+
+ Simple classification
+ 2-3
+ One per category minimum
+
+
+ Complex formatting
+ 3-5
+ Show variations
+
+
+ Nuanced style
+ 4-6
+ Capture full range
+
+
+ Edge cases
+ 1-2
+ Alongside normal examples
+
+
+
+ Symptom
+ Likely Cause
+ Solution
+
+
+ Too long
+ No length constraint
+ Add word/sentence limits
+
+
+ Too short
+ Lacks detail request
+ Ask for elaboration
+
+
+ Off-topic
+ Vague instructions
+ Be more specific
+
+
+ Wrong format
+ Format not specified
+ Define exact structure
+
+
+ Wrong tone
+ Audience not clear
+ Specify audience/style
+
+
+ Inconsistent
+ No examples provided
+ Add few-shot examples
+
+
+
+
Ready to Ship
+
+
Output consistently meets requirements
+
Edge cases are handled appropriately
+
Format is reliable and parseable
+
Further improvements show diminishing returns
+
+
+
+
Keep Iterating
+
+
Output is inconsistent across runs
+
Edge cases cause failures
+
Critical requirements are missed
+
You haven't tested enough variations
+
+
+
+
+
Use JSON When
+
+
Programmatic parsing needed
+
API responses
+
Strict type requirements
+
JavaScript/Web integration
+
Compact representation
+
+
+
+
Use YAML When
+
+
Human readability matters
+
Configuration files
+
Comments are needed
+
DevOps/Infrastructure
+
Deep nested structures
+
+
+
+
+
Do
+
+ - Give a clear identity
+ - List specific capabilities
+ - Set explicit boundaries
+ - Define the tone and style
+ - Include example responses
+
+
+
+
Don't
+
+ - Be vague about the role
+ - Forget to set limits
+ - Make it too long (500 words max)
+ - Contradict yourself
+ - Assume the AI will "figure it out"
+
+
+
+
+
Single Prompt Struggles
+
+
Multi-step reasoning gets confused
+
Different "modes" of thinking clash
+
Complex outputs lack consistency
+
No opportunity for quality control
+
+
+
+
Chaining Solves This
+
+
Each step focuses on one task
+
Specialized prompts for each mode
+
Validate between steps
+
Debug and improve individual steps
+
+
+
+
+
→
+
+
→
+
+
+
Prompt 3
+
(Generate)
+
+
Output
+
+
+
Best For
+
Document summarization, report generation, content repurposing, data-to-narrative conversion
+
+
Best For
+
Code refactoring, project planning, troubleshooting, strategic decision-making, complex problem solving
+
+
Best For
+
Marketing copy, creative writing, email drafts, presentations, any content that benefits from revision
+
+
+
Reducing Latency
+
+
Parallelize independent steps
+
Cache intermediate results
+
Use smaller models for simple steps
+
Batch similar operations
+
+
+
+
Reducing Cost
+
+
Use cheaper models for classification
+
Limit iterations in loops
+
Short-circuit when possible
+
Cache repeated queries
+
+
+
+
Improving Reliability
+
+
Add validation between steps
+
Include retry logic
+
Log intermediate results
+
Implement fallback paths
+
+
+
+
+
Chaining Enables
+
+
Complex multi-step workflows
+
Higher quality through specialization
+
Better error handling and validation
+
Modular, reusable prompt components
+
+
+
+
Key Principles
+
+
Break complex tasks into simple steps
+
Design clear interfaces between steps
+
Validate intermediate outputs
+
Build in error handling and fallbacks
+
Optimize for your constraints
+
+
+
+
How RAG Works:
+
+
+ 1
+ User asks: "What's our refund policy?"
+
+
+ 2
+ System searches your documents for "refund policy"
+
+
+ 3
+ Finds relevant section from your policy document
+
+
+ 4
+ Sends to AI: "Based on this policy: [text], answer: What's our refund policy?"
+
+
+ 5
+ AI generates accurate answer using your actual policy
+
+
+
+
+
RAG Advantages
+
+ - Uses your actual, current data
+ - Reduces hallucinations
+ - Can cite sources
+ - Easy to update (just update documents)
+ - No expensive fine-tuning needed
+
+
+
+
When to Use RAG
+
+ - Customer support bots
+ - Documentation search
+ - Internal knowledge bases
+ - Any domain-specific Q&A
+ - When accuracy matters
+
+
+
+
+
System Prompts
+
Define the agent's identity, capabilities, constraints, and behavioral guidelines. These are the agent's "constitution."
+
+
+
Planning Prompts
+
Guide how agents break down complex goals into actionable steps. Critical for multi-step reasoning.
+
+
+
Tool-Use Prompts
+
Describe available tools and when/how to use them. Agents must understand their capabilities.
+
+
+
Reflection Prompts
+
Enable agents to evaluate their own outputs, catch errors, and improve iteratively.
+
+
+
+
+
→
+
+
→
+
+
→
+
(repeat)
+
+
+
+
Create Plan
+
Break goal into steps
+
+
↓
+
+
+
→
+
+
→
+
+
→
+
...
+
+
↓
+
+
Revise if Needed
+
Adapt plan based on results
+
+
+
+
+
+
Coordinator
+
Manages workflow
+
+
+
⟷
+
↕
+
+
+
+
+
↓
+
+
Planner Agent
+
Breaks down task
+
+
↓
+
+
Researcher Agent
+
Gathers information
+
+
↓
+
+
Writer Agent
+
Creates content
+
+
↓
+
+
Reviewer Agent
+
Quality checks
+
+
↓
+
+
+
+
+
+ | Today |
+ Tomorrow |
+
+
+
+
+ | Writing detailed prompts |
+ Designing agent systems |
+
+
+ | Manual prompt optimization |
+ Automated prompt tuning |
+
+
+ | Single-model expertise |
+ Multi-model orchestration |
+
+
+ | Text-focused interaction |
+ Multimodal fluency |
+
+
+ | Individual productivity |
+ Team-AI collaboration |
+
+
+
+
+
+
+
+
+
+
Agent
+
Autonomous AI system
+
+
+ powered by
+ ↓
+
+
+
+
+
+
+
Skill
+
Reusable expertise
+
+
+
+
+
+
Skill
+
Reusable expertise
+
+
+
+
+
+
Skill
+
Reusable expertise
+
+
+
+ composed of
+ ↓
+
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
Prompts are atoms → Skills are molecules → Agents are complete structures
+
+
+
+
+
+
→
+
+
+
+ ↺ Loop until complete
+
+
+
→
+
+
+
+
+
+
System Prompts
+
Identity & Role
+
+
+
Planning Prompts
+
How to Think
+
+
+
Tool Prompts
+
How to Act
+
+
+
Recovery Prompts
+
How to Recover
+
+
+
+
📄 SKILL.md (Required)
+
The main instruction file. Contains the core expertise, guidelines, and behaviors that define the skill.
+
+
+
📚 Reference Docs
+
Supporting documentation, examples, and context the agent can reference while working.
+
+
+
🔧 Scripts & Tools
+
Helper scripts, templates, or tool configurations that support the skill's functionality.
+
+
+
⚙️ Configuration
+
Settings, parameters, and customization options for adapting the skill to different contexts.
+
+
+
+ 📁
+ code-review-skill/
+
+
+
+ 📄
+ SKILL.md
+ Core review guidelines
+
+
+ 📄
+ security-checklist.md
+ Security patterns
+
+
+ 📄
+ performance-tips.md
+ Optimization guide
+
+
+
+ 📁
+ language-specific/
+
+
+
+ 📄
+ python.md
+ Python best practices
+
+
+ 📄
+ javascript.md
+ JavaScript patterns
+
+
+ 📄
+ rust.md
+ Rust guidelines
+
+
+
+
+
+
+
Simple Prompt
+
+
Single instruction
+
One-off use
+
Limited context
+
Generic approach
+
No supporting materials
+
+
+
+
Skill
+
+
Comprehensive instruction set
+
Reusable across projects
+
Rich context with references
+
Domain-specific expertise
+
Supporting docs, scripts, configs
+
+
+
+
+ Code Review
+
+
+
+
+ Security Audit
+
+
+
+
+ Documentation
+
+
=
+
+ Complete Code Quality Agent
+
+
+
+ AI Agent
+
+
+
+
Code Review
+
Skill 1
+
+
+
+
Test Writing
+
Skill 3
+
+
+
+
+
+
Core Prompts
+
Planning • Tools • Recovery • Memory
+
+
+
+
+
+
Fatih Kadir Akın
+
Creator of Awesome ChatGPT Prompts (a.k.a. prompts.chat)
+
+ Software developer from Istanbul, leading Developer Relations at Teknasyon. Author of books on JavaScript and prompt engineering. Open-source advocate specializing in web technologies and AI-assisted development.
+
+
+
+
+
+ GPT-4o
+ 128K tokens
+
+
+ GPT-5
+ 400K tokens
+
+
+ Claude Sonnet 4
+ 1M tokens
+
+
+ Gemini 2.5
+ 1M tokens
+
+
+ Llama 4
+ 1M-10M tokens
+
+
+ DeepSeek R1
+ 128K tokens
+
+
+
+ Quick questions
+ Faster models like GPT-4o or Claude 3.5 Sonnet
+
+
+ Hard problems
+ Smarter models like GPT-5.2 or Claude 4.5 Opus
+
+
+ Writing code
+ Code-focused models or the smartest general models
+
+
+ Long documents
+ Models with big context windows (Claude, Gemini)
+
+
+ Current events
+ Models with internet access
+
+
+
+ Creation
+ Write, Create, Generate, Compose, Design
+
+
+ Analysis
+ Analyze, Evaluate, Compare, Assess, Review
+
+
+ Transformation
+ Convert, Translate, Reformat, Summarize, Expand
+
+
+ Explanation
+ Explain, Describe, Clarify, Define, Illustrate
+
+
+ Problem-solving
+ Solve, Debug, Fix, Optimize, Improve
+
+
+
+ Audience
+ Who will read/use this?
+
+
+ Length
+ How long/short should it be?
+
+
+ Tone
+ Formal? Casual? Technical?
+
+
+ Format
+ Prose? List? Table? Code?
+
+
+ Scope
+ What to include/exclude?
+
+
+ Purpose
+ What should this accomplish?
+
+
+
+ Too verbose
+ Add "Be concise" or length limits
+
+
+ Too vague
+ Add specific examples or constraints
+
+
+ Wrong format
+ Specify exact output structure
+
+
+ Missing aspects
+ Add "Make sure to include..."
+
+
+ Wrong tone
+ Specify audience and style
+
+
+ Inaccurate
+ Request citations or step-by-step reasoning
+
+
+
+ Code review
+ Senior developer + mentor
+
+
+ Writing feedback
+ Editor + target audience member
+
+
+ Business strategy
+ Consultant + industry expert
+
+
+ Learning new topic
+ Patient teacher + practitioner
+
+
+ Creative writing
+ Specific genre author
+
+
+ Technical explanation
+ Expert + communicator
+
+
+ Problem-solving
+ Domain specialist + generalist
+
+
+
"Let's think step by step."
+
"Think through this carefully."
+
"Work through this problem methodically."
+
"Show your reasoning at each step."
+
"Explain your thought process."
+
+
+
Use CoT For
+
+
**Math problems** — Reduces calculation errors
+
**Logic puzzles** — Prevents skipped steps
+
**Complex analysis** — Organizes thinking
+
**Code debugging** — Traces execution
+
**Decision making** — Weighs trade-offs
+
+
+
+
Skip CoT For
+
+
**Simple Q&A** — Unnecessary overhead
+
**Creative writing** — Can constrain creativity
+
**Factual lookups** — No reasoning needed
+
**Translation** — Direct task
+
**Summarization** — Usually straightforward
+
+
+
+
+ Simple classification
+ 2-3
+ One per category minimum
+
+
+ Complex formatting
+ 3-5
+ Show variations
+
+
+ Nuanced style
+ 4-6
+ Capture full range
+
+
+ Edge cases
+ 1-2
+ Alongside normal examples
+
+
+
+ Symptom
+ Likely Cause
+ Solution
+
+
+ Too long
+ No length constraint
+ Add word/sentence limits
+
+
+ Too short
+ Lacks detail request
+ Ask for elaboration
+
+
+ Off-topic
+ Vague instructions
+ Be more specific
+
+
+ Wrong format
+ Format not specified
+ Define exact structure
+
+
+ Wrong tone
+ Audience not clear
+ Specify audience/style
+
+
+ Inconsistent
+ No examples provided
+ Add few-shot examples
+
+
+
+
Ready to Ship
+
+
Output consistently meets requirements
+
Edge cases are handled appropriately
+
Format is reliable and parseable
+
Further improvements show diminishing returns
+
+
+
+
Keep Iterating
+
+
Output is inconsistent across runs
+
Edge cases cause failures
+
Critical requirements are missed
+
You haven't tested enough variations
+
+
+
+
+
Use JSON When
+
+
Programmatic parsing needed
+
API responses
+
Strict type requirements
+
JavaScript/Web integration
+
Compact representation
+
+
+
+
Use YAML When
+
+
Human readability matters
+
Configuration files
+
Comments are needed
+
DevOps/Infrastructure
+
Deep nested structures
+
+
+
+
+
Do
+
+ - Give a clear identity
+ - List specific capabilities
+ - Set explicit boundaries
+ - Define the tone and style
+ - Include example responses
+
+
+
+
Don't
+
+ - Be vague about the role
+ - Forget to set limits
+ - Make it too long (500 words max)
+ - Contradict yourself
+ - Assume the AI will "figure it out"
+
+
+
+
+
Single Prompt Struggles
+
+
Multi-step reasoning gets confused
+
Different "modes" of thinking clash
+
Complex outputs lack consistency
+
No opportunity for quality control
+
+
+
+
Chaining Solves This
+
+
Each step focuses on one task
+
Specialized prompts for each mode
+
Validate between steps
+
Debug and improve individual steps
+
+
+
+
+
→
+
+
→
+
+
+
Prompt 3
+
(Generate)
+
+
Output
+
+
+
Best For
+
Document summarization, report generation, content repurposing, data-to-narrative conversion
+
+
Best For
+
Code refactoring, project planning, troubleshooting, strategic decision-making, complex problem solving
+
+
Best For
+
Marketing copy, creative writing, email drafts, presentations, any content that benefits from revision
+
+
+
Reducing Latency
+
+
Parallelize independent steps
+
Cache intermediate results
+
Use smaller models for simple steps
+
Batch similar operations
+
+
+
+
Reducing Cost
+
+
Use cheaper models for classification
+
Limit iterations in loops
+
Short-circuit when possible
+
Cache repeated queries
+
+
+
+
Improving Reliability
+
+
Add validation between steps
+
Include retry logic
+
Log intermediate results
+
Implement fallback paths
+
+
+
+
+
Chaining Enables
+
+
Complex multi-step workflows
+
Higher quality through specialization
+
Better error handling and validation
+
Modular, reusable prompt components
+
+
+
+
Key Principles
+
+
Break complex tasks into simple steps
+
Design clear interfaces between steps
+
Validate intermediate outputs
+
Build in error handling and fallbacks
+
Optimize for your constraints
+
+
+
+
How RAG Works:
+
+
+ 1
+ User asks: "What's our refund policy?"
+
+
+ 2
+ System searches your documents for "refund policy"
+
+
+ 3
+ Finds relevant section from your policy document
+
+
+ 4
+ Sends to AI: "Based on this policy: [text], answer: What's our refund policy?"
+
+
+ 5
+ AI generates accurate answer using your actual policy
+
+
+
+
+
RAG Advantages
+
+ - Uses your actual, current data
+ - Reduces hallucinations
+ - Can cite sources
+ - Easy to update (just update documents)
+ - No expensive fine-tuning needed
+
+
+
+
When to Use RAG
+
+ - Customer support bots
+ - Documentation search
+ - Internal knowledge bases
+ - Any domain-specific Q&A
+ - When accuracy matters
+
+
+
+
+
System Prompts
+
Define the agent's identity, capabilities, constraints, and behavioral guidelines. These are the agent's "constitution."
+
+
+
Planning Prompts
+
Guide how agents break down complex goals into actionable steps. Critical for multi-step reasoning.
+
+
+
Tool-Use Prompts
+
Describe available tools and when/how to use them. Agents must understand their capabilities.
+
+
+
Reflection Prompts
+
Enable agents to evaluate their own outputs, catch errors, and improve iteratively.
+
+
+
+
+
→
+
+
→
+
+
→
+
(repeat)
+
+
+
+
Create Plan
+
Break goal into steps
+
+
↓
+
+
+
→
+
+
→
+
+
→
+
...
+
+
↓
+
+
Revise if Needed
+
Adapt plan based on results
+
+
+
+
+
+
Coordinator
+
Manages workflow
+
+
+
⟷
+
↕
+
+
+
+
+
↓
+
+
Planner Agent
+
Breaks down task
+
+
↓
+
+
Researcher Agent
+
Gathers information
+
+
↓
+
+
Writer Agent
+
Creates content
+
+
↓
+
+
Reviewer Agent
+
Quality checks
+
+
↓
+
+
+
+
+
+ | Today |
+ Tomorrow |
+
+
+
+
+ | Writing detailed prompts |
+ Designing agent systems |
+
+
+ | Manual prompt optimization |
+ Automated prompt tuning |
+
+
+ | Single-model expertise |
+ Multi-model orchestration |
+
+
+ | Text-focused interaction |
+ Multimodal fluency |
+
+
+ | Individual productivity |
+ Team-AI collaboration |
+
+
+
+
+
+
+
+
+
+
Agent
+
Autonomous AI system
+
+
+ powered by
+ ↓
+
+
+
+
+
+
+
Skill
+
Reusable expertise
+
+
+
+
+
+
Skill
+
Reusable expertise
+
+
+
+
+
+
Skill
+
Reusable expertise
+
+
+
+ composed of
+ ↓
+
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
Prompts are atoms → Skills are molecules → Agents are complete structures
+
+
+
+
+
+
→
+
+
+
+ ↺ Loop until complete
+
+
+
→
+
+
+
+
+
+
System Prompts
+
Identity & Role
+
+
+
Planning Prompts
+
How to Think
+
+
+
Tool Prompts
+
How to Act
+
+
+
Recovery Prompts
+
How to Recover
+
+
+
+
📄 SKILL.md (Required)
+
The main instruction file. Contains the core expertise, guidelines, and behaviors that define the skill.
+
+
+
📚 Reference Docs
+
Supporting documentation, examples, and context the agent can reference while working.
+
+
+
🔧 Scripts & Tools
+
Helper scripts, templates, or tool configurations that support the skill's functionality.
+
+
+
⚙️ Configuration
+
Settings, parameters, and customization options for adapting the skill to different contexts.
+
+
+
+ 📁
+ code-review-skill/
+
+
+
+ 📄
+ SKILL.md
+ Core review guidelines
+
+
+ 📄
+ security-checklist.md
+ Security patterns
+
+
+ 📄
+ performance-tips.md
+ Optimization guide
+
+
+
+ 📁
+ language-specific/
+
+
+
+ 📄
+ python.md
+ Python best practices
+
+
+ 📄
+ javascript.md
+ JavaScript patterns
+
+
+ 📄
+ rust.md
+ Rust guidelines
+
+
+
+
+
+
+
Simple Prompt
+
+
Single instruction
+
One-off use
+
Limited context
+
Generic approach
+
No supporting materials
+
+
+
+
Skill
+
+
Comprehensive instruction set
+
Reusable across projects
+
Rich context with references
+
Domain-specific expertise
+
Supporting docs, scripts, configs
+
+
+
+
+ Code Review
+
+
+
+
+ Security Audit
+
+
+
+
+ Documentation
+
+
=
+
+ Complete Code Quality Agent
+
+
+
+ AI Agent
+
+
+
+
Code Review
+
Skill 1
+
+
+
+
Test Writing
+
Skill 3
+
+
+
+
+
+
Core Prompts
+
Planning • Tools • Recovery • Memory
+
+
+
+
+
+
Fatih Kadir Akın
+
Creator of Awesome ChatGPT Prompts (a.k.a. prompts.chat)
+
+ Software developer from Istanbul, leading Developer Relations at Teknasyon. Author of books on JavaScript and prompt engineering. Open-source advocate specializing in web technologies and AI-assisted development.
+
+
+
+
+
+ GPT-4o
+ 128K tokens
+
+
+ GPT-5
+ 400K tokens
+
+
+ Claude Sonnet 4
+ 1M tokens
+
+
+ Gemini 2.5
+ 1M tokens
+
+
+ Llama 4
+ 1M-10M tokens
+
+
+ DeepSeek R1
+ 128K tokens
+
+
+
+ Quick questions
+ Faster models like GPT-4o or Claude 3.5 Sonnet
+
+
+ Hard problems
+ Smarter models like GPT-5.2 or Claude 4.5 Opus
+
+
+ Writing code
+ Code-focused models or the smartest general models
+
+
+ Long documents
+ Models with big context windows (Claude, Gemini)
+
+
+ Current events
+ Models with internet access
+
+
+
+ Creation
+ Write, Create, Generate, Compose, Design
+
+
+ Analysis
+ Analyze, Evaluate, Compare, Assess, Review
+
+
+ Transformation
+ Convert, Translate, Reformat, Summarize, Expand
+
+
+ Explanation
+ Explain, Describe, Clarify, Define, Illustrate
+
+
+ Problem-solving
+ Solve, Debug, Fix, Optimize, Improve
+
+
+
+ Audience
+ Who will read/use this?
+
+
+ Length
+ How long/short should it be?
+
+
+ Tone
+ Formal? Casual? Technical?
+
+
+ Format
+ Prose? List? Table? Code?
+
+
+ Scope
+ What to include/exclude?
+
+
+ Purpose
+ What should this accomplish?
+
+
+
+ Too verbose
+ Add "Be concise" or length limits
+
+
+ Too vague
+ Add specific examples or constraints
+
+
+ Wrong format
+ Specify exact output structure
+
+
+ Missing aspects
+ Add "Make sure to include..."
+
+
+ Wrong tone
+ Specify audience and style
+
+
+ Inaccurate
+ Request citations or step-by-step reasoning
+
+
+
+ Code review
+ Senior developer + mentor
+
+
+ Writing feedback
+ Editor + target audience member
+
+
+ Business strategy
+ Consultant + industry expert
+
+
+ Learning new topic
+ Patient teacher + practitioner
+
+
+ Creative writing
+ Specific genre author
+
+
+ Technical explanation
+ Expert + communicator
+
+
+ Problem-solving
+ Domain specialist + generalist
+
+
+
"Let's think step by step."
+
"Think through this carefully."
+
"Work through this problem methodically."
+
"Show your reasoning at each step."
+
"Explain your thought process."
+
+
+
Use CoT For
+
+
**Math problems** — Reduces calculation errors
+
**Logic puzzles** — Prevents skipped steps
+
**Complex analysis** — Organizes thinking
+
**Code debugging** — Traces execution
+
**Decision making** — Weighs trade-offs
+
+
+
+
Skip CoT For
+
+
**Simple Q&A** — Unnecessary overhead
+
**Creative writing** — Can constrain creativity
+
**Factual lookups** — No reasoning needed
+
**Translation** — Direct task
+
**Summarization** — Usually straightforward
+
+
+
+
+ Simple classification
+ 2-3
+ One per category minimum
+
+
+ Complex formatting
+ 3-5
+ Show variations
+
+
+ Nuanced style
+ 4-6
+ Capture full range
+
+
+ Edge cases
+ 1-2
+ Alongside normal examples
+
+
+
+ Symptom
+ Likely Cause
+ Solution
+
+
+ Too long
+ No length constraint
+ Add word/sentence limits
+
+
+ Too short
+ Lacks detail request
+ Ask for elaboration
+
+
+ Off-topic
+ Vague instructions
+ Be more specific
+
+
+ Wrong format
+ Format not specified
+ Define exact structure
+
+
+ Wrong tone
+ Audience not clear
+ Specify audience/style
+
+
+ Inconsistent
+ No examples provided
+ Add few-shot examples
+
+
+
+
Ready to Ship
+
+
Output consistently meets requirements
+
Edge cases are handled appropriately
+
Format is reliable and parseable
+
Further improvements show diminishing returns
+
+
+
+
Keep Iterating
+
+
Output is inconsistent across runs
+
Edge cases cause failures
+
Critical requirements are missed
+
You haven't tested enough variations
+
+
+
+
+
Use JSON When
+
+
Programmatic parsing needed
+
API responses
+
Strict type requirements
+
JavaScript/Web integration
+
Compact representation
+
+
+
+
Use YAML When
+
+
Human readability matters
+
Configuration files
+
Comments are needed
+
DevOps/Infrastructure
+
Deep nested structures
+
+
+
+
+
Do
+
+ - Give a clear identity
+ - List specific capabilities
+ - Set explicit boundaries
+ - Define the tone and style
+ - Include example responses
+
+
+
+
Don't
+
+ - Be vague about the role
+ - Forget to set limits
+ - Make it too long (500 words max)
+ - Contradict yourself
+ - Assume the AI will "figure it out"
+
+
+
+
+
Single Prompt Struggles
+
+
Multi-step reasoning gets confused
+
Different "modes" of thinking clash
+
Complex outputs lack consistency
+
No opportunity for quality control
+
+
+
+
Chaining Solves This
+
+
Each step focuses on one task
+
Specialized prompts for each mode
+
Validate between steps
+
Debug and improve individual steps
+
+
+
+
+
→
+
+
→
+
+
+
Prompt 3
+
(Generate)
+
+
Output
+
+
+
Best For
+
Document summarization, report generation, content repurposing, data-to-narrative conversion
+
+
Best For
+
Code refactoring, project planning, troubleshooting, strategic decision-making, complex problem solving
+
+
Best For
+
Marketing copy, creative writing, email drafts, presentations, any content that benefits from revision
+
+
+
Reducing Latency
+
+
Parallelize independent steps
+
Cache intermediate results
+
Use smaller models for simple steps
+
Batch similar operations
+
+
+
+
Reducing Cost
+
+
Use cheaper models for classification
+
Limit iterations in loops
+
Short-circuit when possible
+
Cache repeated queries
+
+
+
+
Improving Reliability
+
+
Add validation between steps
+
Include retry logic
+
Log intermediate results
+
Implement fallback paths
+
+
+
+
+
Chaining Enables
+
+
Complex multi-step workflows
+
Higher quality through specialization
+
Better error handling and validation
+
Modular, reusable prompt components
+
+
+
+
Key Principles
+
+
Break complex tasks into simple steps
+
Design clear interfaces between steps
+
Validate intermediate outputs
+
Build in error handling and fallbacks
+
Optimize for your constraints
+
+
+
+
How RAG Works:
+
+
+ 1
+ User asks: "What's our refund policy?"
+
+
+ 2
+ System searches your documents for "refund policy"
+
+
+ 3
+ Finds relevant section from your policy document
+
+
+ 4
+ Sends to AI: "Based on this policy: [text], answer: What's our refund policy?"
+
+
+ 5
+ AI generates accurate answer using your actual policy
+
+
+
+
+
RAG Advantages
+
+ - Uses your actual, current data
+ - Reduces hallucinations
+ - Can cite sources
+ - Easy to update (just update documents)
+ - No expensive fine-tuning needed
+
+
+
+
When to Use RAG
+
+ - Customer support bots
+ - Documentation search
+ - Internal knowledge bases
+ - Any domain-specific Q&A
+ - When accuracy matters
+
+
+
+
+
System Prompts
+
Define the agent's identity, capabilities, constraints, and behavioral guidelines. These are the agent's "constitution."
+
+
+
Planning Prompts
+
Guide how agents break down complex goals into actionable steps. Critical for multi-step reasoning.
+
+
+
Tool-Use Prompts
+
Describe available tools and when/how to use them. Agents must understand their capabilities.
+
+
+
Reflection Prompts
+
Enable agents to evaluate their own outputs, catch errors, and improve iteratively.
+
+
+
+
+
→
+
+
→
+
+
→
+
(repeat)
+
+
+
+
Create Plan
+
Break goal into steps
+
+
↓
+
+
+
→
+
+
→
+
+
→
+
...
+
+
↓
+
+
Revise if Needed
+
Adapt plan based on results
+
+
+
+
+
+
Coordinator
+
Manages workflow
+
+
+
⟷
+
↕
+
+
+
+
+
↓
+
+
Planner Agent
+
Breaks down task
+
+
↓
+
+
Researcher Agent
+
Gathers information
+
+
↓
+
+
Writer Agent
+
Creates content
+
+
↓
+
+
Reviewer Agent
+
Quality checks
+
+
↓
+
+
+
+
+
+ | Today |
+ Tomorrow |
+
+
+
+
+ | Writing detailed prompts |
+ Designing agent systems |
+
+
+ | Manual prompt optimization |
+ Automated prompt tuning |
+
+
+ | Single-model expertise |
+ Multi-model orchestration |
+
+
+ | Text-focused interaction |
+ Multimodal fluency |
+
+
+ | Individual productivity |
+ Team-AI collaboration |
+
+
+
+
+
+
+
+
+
+
Agent
+
Autonomous AI system
+
+
+ powered by
+ ↓
+
+
+
+
+
+
+
Skill
+
Reusable expertise
+
+
+
+
+
+
Skill
+
Reusable expertise
+
+
+
+
+
+
Skill
+
Reusable expertise
+
+
+
+ composed of
+ ↓
+
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
+ Prompt
+
+
+
Prompts are atoms → Skills are molecules → Agents are complete structures
+
+
+
+
+
+
→
+
+
+
+ ↺ Loop until complete
+
+
+
→
+
+
+
+
+
+
System Prompts
+
Identity & Role
+
+
+
Planning Prompts
+
How to Think
+
+
+
Tool Prompts
+
How to Act
+
+
+
Recovery Prompts
+
How to Recover
+
+
+
+
📄 SKILL.md (Required)
+
The main instruction file. Contains the core expertise, guidelines, and behaviors that define the skill.
+
+
+
📚 Reference Docs
+
Supporting documentation, examples, and context the agent can reference while working.
+
+
+
🔧 Scripts & Tools
+
Helper scripts, templates, or tool configurations that support the skill's functionality.
+
+
+
⚙️ Configuration
+
Settings, parameters, and customization options for adapting the skill to different contexts.
+
+
+
+ 📁
+ code-review-skill/
+
+
+
+ 📄
+ SKILL.md
+ Core review guidelines
+
+
+ 📄
+ security-checklist.md
+ Security patterns
+
+
+ 📄
+ performance-tips.md
+ Optimization guide
+
+
+
+ 📁
+ language-specific/
+
+
+
+ 📄
+ python.md
+ Python best practices
+
+
+ 📄
+ javascript.md
+ JavaScript patterns
+
+
+ 📄
+ rust.md
+ Rust guidelines
+
+
+
+
+
+
+
Simple Prompt
+
+
Single instruction
+
One-off use
+
Limited context
+
Generic approach
+
No supporting materials
+
+
+
+
Skill
+
+
Comprehensive instruction set
+
Reusable across projects
+
Rich context with references
+
Domain-specific expertise
+
Supporting docs, scripts, configs
+
+
+
+
+ Code Review
+
+
+
+
+ Security Audit
+
+
+
+
+ Documentation
+
+
=
+
+ Complete Code Quality Agent
+
+
+
+ AI Agent
+
+
+
+
Code Review
+
Skill 1
+
+
+
+
Test Writing
+
Skill 3
+
+
+
+
+
+
Core Prompts
+
Planning • Tools • Recovery • Memory
+
+
+