A free Claude skill that turns one messy sentence into a finished prompt. But it doesn't remember your decisions in a new chat on its own

The repo is MIT licensed and completely free, with over 13k stars on GitHub. If you'd rather go straight there, the button opens the repo; that's where you download the ZIP.

What's extra below: what the skill does and doesn't do, the sentence pattern that triggers it, and how the feature the guide calls "remembers your decisions in every new chat" actually works.

The problem is familiar: you write a rough prompt, Claude does exactly what you typed, and you spend three more messages explaining what you actually meant. Prompt Master cuts that loop at the start. From your sentence it pulls out the task, the output format, the limits and the finish condition; if something critical is missing it asks at most three questions, and at the end it gives you one copyable prompt.

The last step matters: the skill doesn't do the job itself, it writes the prompt that will. So one message goes to getting the prompt written. The gain is that the "no, I meant this" messages that would have followed disappear.

This is the route the repo recommends, no terminal needed.

If you use Claude Code, you clone the repo straight into the skills folder. The repo lists this as the second option; it works, you just have to update it yourself with git pull.

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

git clone https://github.com/nidhinjs/prompt-master.git "$HOME\.claude\skills\prompt-master"

After restarting Claude Code you can call it directly by typing /prompt-master.

The guide's line "stop writing careful prompts, write messy ones" is half true. The skill's own description is clear: it only kicks in when you ask it to write, fix or adapt a prompt for a specific tool. A plain "build a landing page" message triggers nothing, and Claude just starts on the job. Your sentence has to say you want a prompt and which tool it's for.

Write a Claude Code prompt: a landing page for my agency with a booking section

Write a Cursor prompt: refactor my auth module, touch no other files

Fix this prompt for ChatGPT: [paste your prompt]

Write a Midjourney prompt: a realistic samurai standing in the rain at night

The prompt that comes out is shaped for the target tool. For coding agents it adds a starting state, the files that may be touched, an "ask before doing this" list and a finish condition; for image tools, comma-separated descriptors and a negative prompt. Read it, tweak if needed, then run it.

The English guide says "it remembers the decisions you make, every new chat starts already knowing your stack." The skill's own file says something else. Prompt Master has no persistent memory. What it does: it collects decisions already made in that chat, your stack, naming conventions, the things you said not to use, and puts them as a memory block at the top of the prompt it writes. Within the same chat that's very useful, because the new prompt doesn't contradict earlier decisions.

## Memory (carried forward from earlier)
- Stack: React 18 + TypeScript + Supabase
- Auth via JWT in httpOnly cookies, not localStorage
- Component names in PascalCase, no default exports
- Tailwind only, no separate CSS files

If you want it to persist, write this list once and put it in a Project's instructions in Claude; every chat you open in that project sees it. In Claude Code the CLAUDE.md file in your project folder does the same job, covered separately in 114.

The guide says 35, the repo's current version has 37 patterns. You don't need to memorize them, the skill checks them anyway. But knowing the few that burn the most messages fixes the prompts you write without it too.

The skill runs inside Claude; it doesn't install into ChatGPT or Cursor. The "works there too" part means this: in Claude you can have it write prompts for ChatGPT, Cursor, Gemini, Midjourney and more than thirty other tools, then paste the result into that tool. Name the target tool in your sentence and the prompt follows that tool's rules.