prompt-master

---
name: prompt-master
description: Use when writing, fixing, improving, or adapting a prompt for a specific AI tool — extracts intent across 9 dimensions, routes to per-model best practices, and outputs one paste-ready prompt with zero wasted tokens.
source: https://github.com/nidhinjs/prompt-master
generated: 2026-07-02T18:43:06.331Z
category: tool
audience: general
---

## When to use

- Turning a rough idea into a single production-ready prompt for a named tool (Claude, GPT, Gemini, Midjourney, image/video AI, coding agents)
- Fixing or improving an underperforming prompt that wastes tokens or drifts off-format
- Adapting one prompt across different models that need different prompting styles
- Choosing the right technique (role assignment, few-shot, grounding, chain-of-thought) for the target model
- Avoiding model-specific anti-patterns, e.g. not adding chain-of-thought to reasoning-native models

## Key concepts

### Confirm the target tool first

Never output a prompt without knowing which AI system receives it — different models need different prompting. Ask if ambiguous, but never more than 3 clarifying questions before producing a prompt.

### 9-dimension intent extraction

Silently extract Task, Target tool, Output format, Constraints, Input, Context, Audience, Success criteria, and Examples. Missing critical dimensions (task/tool/format are always critical) trigger clarifying questions.

### Per-model tool routing

Route to model-specific guidance: Claude follows literally and over-engineers (add 'only what's asked'); reasoning models (o3, Qwen3-thinking) want short prompts with NO chain-of-thought; open-weight models need flat, explicit structure.

### Prefer simple techniques

Favor role assignment, few-shot, grounding anchors, and chain-of-thought over complex meta-frameworks (Mixture of Experts, Tree/Graph of Thought) which carry higher fabrication risk in single-prompt contexts.

### Locked output format

Output one copyable prompt block, then a one-line '🎯 Target / 💡 what was optimized', and a short setup note only if genuinely needed. No framework names, no unrequested explanation, build prompts one at a time.

## API reference

```
Output: copyable prompt block + 🎯 Target + 💡 what was optimized
```

The fixed deliverable shape — a paste-ready prompt and a single line of rationale, with fillable placeholders only where relevant.

```
<prompt ready to paste>
🎯 Target: Claude Opus 4.x  💡 Added explicit output format + 'only what's asked' guard to stop over-engineering.
Placeholders where useful: [TONE], [AUDIENCE], [BRAND VOICE], [PRODUCT NAME]
```

```
Model routing cheatsheet
```

Tailor the prompt to the receiving model's strengths and failure modes.

```
Claude 4.x: explicit + literal; XML tags for multi-section; add 'don't refactor beyond what's asked'.
o3/o4-mini/Qwen3-thinking: short clean instructions, NO CoT, <200-word system prompts.
Gemini: format locks + 'cite only certain sources'.
Local (Ollama/Llama/Mistral): flat structure, ask which model is running.
```

## Gotchas

- Do not output a prompt without first confirming the target tool — ask if ambiguous.
- Do not add chain-of-thought or 'think step by step' to reasoning-native models (o3, o4-mini, DeepSeek-R1, Qwen3 thinking) — it degrades output.
- Avoid Mixture of Experts, Tree of Thought, and Graph of Thought in single-prompt contexts unless explicitly requested — higher fabrication risk.
- Do not ask more than 3 clarifying questions before producing a prompt.
- Do not pad the output with explanations the user didn't request, and never show framework names in the output.
- Only activates for explicit prompt-engineering requests — not general conversation, coding, or document writing.

---
Generated by SkillMake from https://github.com/nidhinjs/prompt-master on 2026-07-02T18:43:06.331Z.
Verify against source before relying on details.

File: ~/.claude/skills/prompt-master/SKILL.md