Prompt Master

---
name: prompt-master-skill
description: Claude skill that generates accurate, token-efficient prompts for any AI tool — Claude, GPT, Midjourney, Cursor, and 30+ more.
triggers:
  - write me a prompt for
  - generate a prompt for
  - help me prompt
  - fix my prompt
  - create a midjourney prompt
  - write a cursor prompt
  - make a prompt for claude code
  - I need a better prompt for
---

# Prompt Master

> Skill by [ara.so](https://ara.so) — Daily 2026 Skills collection.

Prompt Master is a Claude skill that writes sharp, token-efficient prompts for any AI tool. It eliminates the re-prompting loop by extracting full intent on the first pass, routing to the right prompt architecture, and auditing every word for necessity before delivery.

**Supports:** Claude, ChatGPT, Gemini, o1/o3/o4, DeepSeek, MiniMax, Qwen, Cursor, Windsurf, Claude Code, GitHub Copilot, Bolt, v0, Lovable, Devin, Perplexity, Midjourney, DALL-E, Stable Diffusion, ComfyUI, Sora, Runway, ElevenLabs, Zapier, Make, and any unknown tool via universal fingerprint.

---

## Installation

### Recommended — Claude.ai (browser)

1. Download the repo as a ZIP from [github.com/nidhinjs/prompt-master](https://github.com/nidhinjs/prompt-master)
2. Go to **claude.ai → Sidebar → Customize → Skills → Upload a Skill**

### Claude Code (CLI)

```bash
mkdir -p ~/.claude/skills
git clone https://github.com/nidhinjs/prompt-master.git ~/.claude/skills/prompt-master

Once installed, the skill activates automatically when Claude detects prompt-generation intent, or you can invoke it explicitly.

---

How to Invoke

Natural language (auto-detected)

Write me a prompt for Cursor to refactor my auth module

I need a prompt for Claude Code to build a REST API

Here's a bad prompt I wrote for GPT-4o, fix it: [paste prompt]

Generate a Midjourney prompt for a cyberpunk city at night

Explicit invocation

/prompt-master

I want to ask Claude Code to build a todo app with React and Supabase

---

What Happens Internally (The Pipeline)

Prompt Master runs a 7-step pipeline on every request. You never see the machinery — only the output.

Step	What It Does
1. Tool Detection	Identifies the target AI system and routes to its profile
2. Intent Extraction	Pulls 9 dimensions: task, input, output, constraints, context, audience, memory, success criteria, examples
3. Clarifying Questions	Asks ≤3 targeted questions if critical info is missing
4. Framework Routing	Selects the right prompt architecture from 12 templates
5. Safe Techniques	Applies role assignment, few-shot, XML structure, grounding anchors, memory blocks
6. Token Efficiency Audit	Strips every word that doesn't change the output
7. Delivery	One clean copyable block + one-line strategy note

---

Prompt Templates (Auto-Selected)

You never choose a template — Prompt Master picks it. Shown here for understanding:

Template	Used For
RTF (Role, Task, Format)	Fast one-shot tasks
CO-STAR	Professional documents, business writing
RISEN	Complex multi-step projects
CRISPE	Creative work, brand voice, iterative content
Chain of Thought	Math, logic, debugging, multi-step analysis
Few-Shot	Consistent structured output, pattern replication
File-Scope Template	Cursor, Windsurf, Copilot — code editing AI
ReAct + Stop Conditions	Claude Code, Devin — autonomous agents
Visual Descriptor	Midjourney, Stable Diffusion, image AI
Prose Description	DALL-E 3
Node Split	ComfyUI (positive/negative node separation)
Universal Fingerprint	Any unknown tool — 4 questions, then generates

---

Real Usage Examples

Example 1 — Image prompt (Midjourney)

You say:

Write me a midjourney prompt for a realistic samurai standing in the rain at night

Prompt Master delivers:

lone samurai standing in heavy rain at night, traditional armor,
neon reflections on wet cobblestone street, cinematic lighting,
dramatic shadows, fog, ultra detailed, photorealistic,
shallow depth of field --ar 16:9 --v 6 --style raw

negative: blurry, low quality, watermark, cartoon, anime, extra limbs

🎯 Target: Midjourney · ⚡ Framework: Visual Descriptor · 💰 Tokens: Light (~60) 💡 Strategy: Comma-separated descriptors over prose, lighting anchored early, version locked, negative prompt prevents style drift.

---

Example 2 — Coding prompt (Claude Code)

You say:

Build a claude code prompt for a landing page that looks like Notion

Prompt Master delivers:

Objective:
Build a fully detailed, production-quality static landing page for a
business dashboard product. Design must match Notion's exact aesthetic —
clean, minimal, typographic, with smooth animations throughout.

Stack:
Single HTML file with embedded CSS and vanilla JS. No frameworks.
No external dependencies except Google Fonts (Inter).

Design Spec — match these exactly:
- Background: #ffffff
- Primary text: #1a1a1a
- Secondary text: #6b7280
- Border color: #e5e7eb
- Accent: #000000
- Font: Inter, weights 400 / 500 / 600 / 700
- Base spacing unit: 8px
- Border radius: 6px on cards, 4px on buttons
- Box shadows: 0 1px 3px rgba(0,0,0,0.08) only

Sections to build in order:
1. Navbar — logo left, nav links center, CTA right. Sticky, border-bottom on scroll.
2. Hero — large left-aligned heading, subtext, two CTAs, mockup placeholder.
3. Logos bar — "Trusted by teams at..." with 5 placeholder names, muted gray.
4. Features — 3-column grid, icon + bold title + 2-line description.
   Hover: border darkens to #d1d5db, translateY(-2px), 200ms ease.
5. How it works — alternating left/right, 3 steps.
6. Pricing — 2 tiers. Free: outlined. Pro: black bg, white text, badge.
7. CTA banner — full-width dark section, centered heading, single CTA.
8. Footer — 4-column links, copyright.

Animations:
- Fade-in + translateY(20px→0) on scroll via IntersectionObserver. Threshold 0.15. 500ms ease-out.
- Feature cards: 100ms stagger delay.
- All hover transitions: 200ms ease.
- Smooth scroll on all anchor links.

Constraints:
- Single file only. All CSS and JS embedded.
- No dependencies except Google Fonts CDN.
- Placeholder images = styled gray divs.
- Pixel-perfect at 375px and 1440px.
- Zero console errors.

Done When:
All 8 sections render, animations trigger without layout shift,
hover states work, navbar sticks and border appears on scroll.

🎯 Target: Claude Code · 💰 Tokens: Medium (~380) 💡 Strategy: Every vague aesthetic cue translated into exact hex values and pixel specs — Claude Code cannot guess wrong.

---

Example 3 — Fix a bad prompt

You say:

Here's a bad prompt I wrote for GPT-4o, fix it:

"write me some python code to handle user auth"

Prompt Master delivers:

You are a senior Python backend engineer.

Task: Implement a complete user authentication module.

Requirements:
- Framework: FastAPI
- Auth method: JWT (access token 15min, refresh token 7 days)
- Password hashing: bcrypt via passlib
- Endpoints: POST /auth/register, POST /auth/login, POST /auth/refresh, POST /auth/logout
- Store refresh tokens in Redis with TTL
- Return format: { access_token, refresh_token, token_type }

Constraints:
- No third-party auth services (Auth0, Firebase, etc.)
- No ORM — use raw asyncpg queries
- Include input validation with Pydantic v2 models
- Include error handling with proper HTTP status codes

Deliver:
- auth/router.py
- auth/schemas.py
- auth/utils.py
- auth/dependencies.py

Do not scaffold a full app. Auth module only.

🎯 Target: GPT-4o · ⚡ Framework: RTF + File-Scope 💡 Strategy: Vague "some python code" collapsed into exact stack, endpoints, constraints, and file deliverables.

---

Example 4 — Stable Diffusion

You say:

Write a stable diffusion prompt for a futuristic city interior, warm lighting

Prompt Master delivers:

Positive:
futuristic city interior, (warm golden lighting:1.3), (volumetric light rays:1.2),
glass and steel architecture, holographic displays, busy atrium,
people walking, ultra detailed, 8k, cinematic composition,
(photorealistic:1.4), depth of field, bokeh background

Negative:
(blurry:1.3), (low quality:1.3), (watermark:1.2), cartoon, anime,
flat lighting, oversaturated, extra limbs, bad anatomy, deformed,
ugly, duplicate, morbid, mutilated, out of frame

🎯 Target: Stable Diffusion · ⚡ Framework: Node Split 💡 Strategy: Weight syntax applied to critical elements, mandatory negative prompt, photorealism locked.

---

Example 5 — Workflow automation (Zapier)

You say:

Write a prompt for Zapier to automatically save Gmail attachments to Google Drive

Prompt Master delivers:

Trigger:
- App: Gmail
- Event: New Email Matching Search
- Search filter: has:attachment

Filter (add a Filter step):
- Only continue if: Attachment exists = True

Action 1:
- App: Gmail
- Action: Get Email
- Field: Use email ID from trigger step

Action 2:
- App: Google Drive
- Action: Upload File
- File: Attachment from Action 1
- Drive: My Drive
- Folder: /Gmail Attachments/{{trigger.from_email}}/{{trigger.date_formatted}}
- File name: {{trigger.subject}} - {{attachment.filename}}

Notes:
- Add a second Action 2 loop if email can have multiple attachments
- Set folder path dynamically using sender email and date to avoid clutter
- Test with a single known email before enabling live Zap

🎯 Target: Zapier · ⚡ Framework: Trigger-Action Map 💡 Strategy: Trigger app + event + filter + field mapping fully specified — no Zapier configuration guesswork.

---

Tool Profile Reference

Prompt Master includes built-in profiles that know the quirks of each tool:

Claude        → XML structure, length spec, no padding
ChatGPT       → Output contract, verbosity control, completion criteria
Gemini        → Grounding anchors, citation rules, format locks
o3/o4-mini    → Short clean instructions only — never add CoT (they think internally)
DeepSeek-R1   → Short instructions, suppresses thinking output if needed
MiniMax       → Temperature hints, thinking tag control
Ollama        → Asks which model is loaded, includes system prompt for Modelfile
Cursor        → File path, function name, do-not-touch list
Claude Code   → Stop conditions, file scope, checkpoint output
Copilot       → Exact function contract as docstring
Bolt/v0       → Stack spec, version, what NOT to scaffold
Devin         → Starting state, target state, stop conditions
Midjourney    → Comma descriptors, --parameters, negative prompts
DALL-E 3      → Prose description, text exclusion, edit vs generate detection
Stable Diff   → Weight syntax (word:1.3), CFG, mandatory negatives
ComfyUI       → Positive/negative node split, checkpoint-specific syntax
Sora/Runway   → Camera movement, duration, cut style
ElevenLabs    → Emotion, pacing, emphasis, speech rate
Zapier/Make   → Trigger + event + action + field mapping

For any unlisted tool, Prompt Master uses the Universal Fingerprint — 4 questions to characterize the tool and generate a quality prompt anyway.

---

Clarifying Questions Logic

Prompt Master asks ≤3 questions only when critical info is genuinely missing. It never asks for information it can reasonably infer.

It will ask when:

• Target tool is ambiguous and it changes the output significantly
• Task scope is completely undefined (e.g. "write a prompt for building an app" — what app?)
• Output format has high-stakes options (e.g. single file vs multi-file for Claude Code)

It will NOT ask:

• For style preferences it can default to sensibly
• For confirmation of things clearly stated
• More than 3 questions ever

---

Key Principles

The core insight

"The best prompt is not the longest. It's the one where every word is load-bearing."

What Prompt Master does differently from prompt generators

• Other tools: Make prompts longer
• Prompt Master: Makes prompts sharper — token efficiency audit strips everything that doesn't change the output

The 9 dimensions of intent it extracts

1. Task — what action needs to happen
2. Input — what the AI is working with
3. Output — what the result should look like
4. Constraints — what must not happen
5. Context — background the AI needs
6. Audience — who the output is for
7. Memory — what should persist across turns
8. Success criteria — how to know it's done
9. Examples — patterns to replicate or avoid

---

Troubleshooting

Prompt Master isn't activating

• Ensure the skill is uploaded/installed correctly in Claude's skill directory
• Try explicit invocation: /prompt-master [your request]
• Make sure your request mentions a target tool or prompt-related intent

Generated prompt is too long for my use case

• Tell Prompt Master the token budget: "Write me a Midjourney prompt, keep it under 50 tokens"
• It will re-run the token efficiency audit with a hard cap

Generated prompt doesn't match my tool's syntax

• Name the exact tool version: "for Stable Diffusion 1.5" vs "for SDXL" — these have different optimal syntax
• If it's an obscure or new tool, paste 1-2 example prompts that work: "Here's a prompt that worked for [Tool X], use that style"

It asked me questions but I want a prompt now

• Say: "Just generate your best guess, I'll refine after"
• Prompt Master will use reasonable defaults and flag its assumptions in the strategy note

The prompt works but I want a variation

• Say: "Give me 3 variations of that prompt, different approaches"
• Or: "Make it more cinematic" / "Make it more technical" — it retains full context

I want to see which framework it used

• Ask: "Which prompt framework did you use and why?"
• It will explain the routing decision

---

Tips for Best Results

# Be specific about the target tool
"Write a prompt for Claude Code"        ✅
"Write a prompt for an AI"              ❌ (will ask for clarification)

# Include your stack when relevant
"Claude Code prompt for a React app using Supabase and Tailwind"   ✅
"Claude Code prompt for an app"                                     ❌

# Paste bad prompts for fixing — Prompt Master thrives on this
"Fix this: write me some code for auth"   ✅

# Give reference style when you have one
"Generate a Midjourney prompt like this example: [paste]"   ✅

# Specify constraints upfront
"Single file only, no external dependencies"   ✅ (saves a clarifying question round)

---

Project Info

• Repo: github.com/nidhinjs/prompt-master
• License: MIT
• Topics: claude-ai, claude-skills, llm, prompt-engineering
• Works with: Claude.ai browser skill system, Claude Code skill directory