Expert prompt engineering skill for Claude Code. Creates optimized, context-aware prompts for any task, any model, any use case.
Built from official Anthropic sources — not academic frameworks, not guesswork.
Because sometimes you need to leave. You're deep in a coding session, you hit a wall — missing domain knowledge, need a legal review, need market research, need to analyze a dataset. Your current AI doesn't have that context, or you need a different tool for the job.
Instead of context-switching, losing focus, and manually writing a prompt for Gemini, Perplexity, ChatGPT, or another LLM — you just say "write me a prompt for deep research on X". The skill:
- Reads your current context — what you're building, what stack you're using, what problem you're solving
- Generates an optimized prompt — with the right structure, real variables from your project, and the right format for the target model
- You paste it into another tool — and get back exactly what you need, without losing your flow
Common use cases:
- You're building a fintech app → need a deep research prompt about compliance regulations → send to Perplexity
- You're writing an API → need a prompt to generate test data that matches your schema → send to GPT-4
- You're debugging → need a prompt to analyze logs with specific patterns → send to Gemini with your files
- You need a system prompt for a chatbot you're building → generate it right here, already matching your codebase
TL;DR: It's a prompt factory that understands your project context and produces ready-to-use prompts for any AI tool — so you never break your flow.
You say "write me a prompt for..." — it automatically gathers context from your current session, then builds a production-ready prompt using a structured 10-part framework derived from Anthropic's own documentation.
write a prompt for a customer support chatbot
A complete, structured prompt with:
- Role and persona definition (auto-detected from your project domain)
- Task context and rules
- 3-5 few-shot examples (using real data from your session when available)
- Input/output XML structure
- Edge case handling
- Fallback behavior
- Output format (matched to your downstream needs)
- Model recommendation
- Tuning tips
| Feature | Description |
|---|---|
| Context-Aware (Phase 0) | Automatically scans session, project files, and conversation to infer model, stack, persona, and format |
| 10-Part Framework | Systematic prompt structure from Anthropic's interactive tutorial |
| 6 Task Templates | Classification, Extraction, Creative, Chatbot, Code Gen, Analysis/RAG |
| Multi-Model | Claude 4.x, GPT-5.1/4o, Gemini 2.5, Llama 4, DeepSeek R1, Mistral, Qwen 3, Grok |
| Advanced Techniques | Reflexion, ReAct, Tree of Thoughts, Skeleton-of-Thought, Emotion Prompting, Self-Consistency, Directional Stimulus |
| Agentic Patterns | Plan-then-Execute, Agent Memory Management, Multi-Agent Decision Protocols, Error Recovery |
| Meta-Prompting | Contrastive Learning (LCP) optimization, Meta-Expert orchestration |
| Prompt Security | Sandwich Defense, Salted XML Tags, Attack Short-Circuiting |
| Long-Context | Lost-in-the-Middle mitigation, MapReduce summarization |
| Multimodal | Temporal grounding, resolution control, OCR-Vision hybrid, Visual Chain-of-Thought |
| Evaluation | LLM-as-Judge scoring, RAG Triad quality checks |
| Prompt Optimization | Compression (MetaGlyph, LLMLingua), caching strategies, model routing/cascading |
| Prompt Debugging | Bisection debugging (O(log n)), eval-driven development (Promptfoo) |
| Domain Patterns | Legal (clause tagging), Medical (NER+LLM), Financial (reasoning/calculation split) |
| Production Patterns | Prompt registries, CI/CD quality gates, observability, fine-tune vs. prompt decision |
| Quality Checklist | 16-point verification before every prompt delivery |
| Anti-Patterns Table | 15 common mistakes with fixes |
| Multilingual Triggers | Activates in any language the user writes in |
Unlike static prompt generators, this skill reads your environment before building:
Session Context
├── Project type (package.json, requirements.txt, go.mod, etc.)
├── Tech stack (dependencies, framework configs)
├── Current task (conversation history, recent tool calls)
├── Target platform (API, chatbot, n8n, Telegram, web app)
├── User's language (matches prompt language automatically)
├── Existing prompts (CLAUDE.md, .cursorrules — avoids conflicts)
└── Real data (recent file reads, API responses → used as examples)
Result: Instead of generic placeholders, you get prompts with real variable names, matching coding style, and domain-appropriate personas.
This skill is synthesized from 4 official Anthropic documents, enhanced with 69-source academic research and 86-source practitioner research (social media, dev blogs, production case studies):
- Prompt Engineering Interactive Tutorial — 9 chapters + 3 appendices
- Claude Prompting Best Practices — production patterns for Claude 4.x
- Prompting Tools — prompt generator, improver, templates
- Prompt Engineering Overview — evaluation-first approach
mkdir -p ~/.claude/skills/claude-prompt-engineering && curl -fsSL https://raw.githubusercontent.com/MOZARTINOS/claude-prompt-engineering/main/claude-prompt-engineering/SKILL.md -o ~/.claude/skills/claude-prompt-engineering/SKILL.md-
Clone this repo:
git clone https://github.com/MOZARTINOS/claude-prompt-engineering.git
-
Copy the skill to your Claude Code skills directory:
cp -r claude-prompt-engineering/claude-prompt-engineering ~/.claude/skills/ -
Restart Claude Code. The skill auto-activates on relevant prompts.
ls ~/.claude/skills/claude-prompt-engineering/SKILL.mdThe skill activates automatically when you ask Claude Code to create a prompt. Trigger phrases include:
| Trigger | Examples |
|---|---|
| English | "write a prompt for...", "create a prompt", "system prompt for...", "prompt template", "LLM prompt", "chatbot instructions" |
| Any language | The skill detects prompt-creation intent regardless of language |
/claude-prompt-engineering
Create a system prompt for a code review bot that checks Python code for security vulnerabilities
Write a prompt for classifying customer emails into categories: billing, technical, feature request, complaint
Build a RAG prompt that answers questions based on uploaded PDF documents
Not every prompt needs all 10 parts. Start with all, then trim to the minimum effective set.
CORE (10-Part Framework)
0. Context Gathering — Auto-detect from session (Phase 0)
1. Role / Persona — Who is the AI?
2. Task Context — Why does this task exist?
3. Tone / Style — How should it communicate?
4. Instructions & Rules — Step-by-step behavior, edge cases
5. Examples (Few-Shot) — THE most effective technique
6. Input Data — Variable content in XML tags
7. Task Reiteration — Restate task near the end (for long prompts)
8. Thinking / Reasoning — "Think step by step" before answering
9. Output Format — Exact structure specification
10. Structured Output — Model-specific enforcement (API level)
ADVANCED (from 69-source deep research)
11. Reflexion (RSIP) — Self-critique before finalizing
12. ReAct — Thought → Action → Observation loops
13. Skeleton-of-Thought — Outline first, expand second
14. Emotion Prompting — Psychological cues (+8-10% fidelity)
15. Directional Stimulus — Keyword anchoring for focus
16. Self-Consistency — Multiple paths, majority vote
17. Tree of Thoughts — Simulated expert panel with backtracking
META & SECURITY
18. Contrastive Optimization — Learn from failure analysis (LCP)
19. Meta-Expert Orchestration— Conductor-Expert panel simulation
20. Sandwich Defense — Repeat constraints after untrusted input
21. Salted XML Tags — Anti-injection with random suffixes
22. LLM-as-Judge — Automated quality scoring
23. RAG Triad — Context/Groundedness/Answer relevance
AGENTIC & LONG-CONTEXT (from 86-source practitioner research)
24. Plan-then-Execute — Secure planning/execution separation
25. Agent Memory Management — 4 memory patterns for agentic systems
26. Multi-Agent Decisions — Consensus vs. voting protocols
27. Agentic Error Recovery — Retry, self-critique, graceful degradation
28. Lost-in-the-Middle — U-shaped attention curve mitigations
29. MapReduce Summarization — Chunk → Map → Reduce for long docs
MULTIMODAL & OPTIMIZATION
30. OCR-Vision Hybrid — Image + OCR text for max accuracy
31. Visual Chain-of-Thought — See → Think → Confirm → Answer
32. Prompt Compression — MetaGlyph, LLMLingua, rule-based
33. Prompt Caching — Provider-specific cache strategies
34. Model Routing — Cost-optimized model cascading
DEBUGGING & PRODUCTION
35. Prompt Bisection — O(log n) systematic debugging
36. Eval-Driven Development — Test-driven prompt CI/CD
| Model | Key Adjustments |
|---|---|
| Claude 4.x | No prefill (4.6+), "think" tool, adaptive thinking, structured outputs |
| GPT-5/5.1 | Hard vs. soft constraints, tool preambles, agentic eagerness calibration |
| Gemini 2.5 | Separate system instructions, Google Search grounding, 1M-token context |
| DeepSeek R1 | Zero-shot > few-shot, minimal prompts, user role preferred |
| Grok (xAI) | Confidence labeling, surface-specific behavior, explicit boundaries |
| Local (Llama 4, Qwen 3, Mistral) | Official instruct templates, more examples (5-10), explicit JSON schema |
| Feature | claude-prompt-engineering | prompt-architect |
|---|---|---|
| Approach | Creates from scratch + context-aware | Transforms existing prompts |
| Source | Anthropic docs + 69 academic + 86 practitioner sources | Academic frameworks (CO-STAR, RISEN) |
| Techniques | 36 (core + advanced + agentic + optimization + debugging) | 7 frameworks |
| Context-aware | Auto-detects project, stack, task | No context gathering |
| Multi-model | Claude 4.x, GPT-5.1, Gemini 2.5, Llama 4, DeepSeek R1, Qwen 3, Grok | Claude only |
| Task templates | 6 ready-to-use | None |
| Security | Sandwich, Salted XML, Short-Circuiting | Not included |
| Multimodal | Video/audio temporal grounding | Not included |
| Evaluation | LLM-as-Judge, RAG Triad | 5-dimension analysis |
| Anti-patterns | 12 documented with fixes | Not included |
| Quality checklist | 13-point verification | Not included |
PRs welcome. If you want to:
- Add templates for new task types
- Add support for more models
- Add trigger phrases in other languages
- Fix outdated model-specific advice
- Improve context detection heuristics
Please open an issue first to discuss the change.