prompt-craft

Skills that turn weak prompts into good ones and stop you wasting tokens.

Install

Claude Code (plugin):

claude plugin marketplace add mustansirlokhandwala/prompt-craft

claude plugin install prompt-craft@prompt-craft

Claude Code (manual): copy the folders in skills/ into ~/.claude/skills/.

Claude.ai: zip a skill folder (for example skills/prompt-improver) and upload it under Settings → Capabilities → Skills.

Use

Just ask:

• “Improve this prompt: write me a blog post about AI”
• “/prompt-improver can u fix my code its broken”
• “How many tokens will this take?” (with files attached)
• “I have an idea for an app but I can’t explain it”
• “Make me a landing page for my bakery” or “my slides look amateur”
• “Reply to this client: sorry we’re late, can deliver Thursday”

The token-budget skill also warns on its own when you paste a lot of text or start a big multi-step task.

Always on

Skills load only when a request matches them. To apply prompt-craft to every prompt, load the short rules in rules/always-on.md (~250 tokens) into the agent’s standing context:

• Claude Code plugin: automatic. A SessionStart hook loads the rules in every session.
• Claude Code manual: append the rules to ~/.claude/CLAUDE.md.
• Claude.ai: paste the rules into your profile preferences or a project’s instructions.
• Antigravity: append the rules to ~/.gemini/GEMINI.md, or copy them into .agents/rules/ and set the rule to Always On.

Step-by-step setup for each tool is in docs/always-on.md.

Measure files from the command line:

python3 skills/token-budget/scripts/estimate_tokens.py notes.md src/*.py --budget 20000

The five habits these skills teach

1. Say what you want, not what to do first. State the goal and what “done” looks like.
2. Give the context only you have. The model cannot know your audience, data or constraints.
3. Define the output. Format, length and what to leave out.
4. Cut the filler. Politeness and repetition cost tokens and change nothing.
5. Allow “I don’t know”. Grounded, honest answers beat confident guesses.

Contributing

New skills are welcome. Each skill is a folder in skills/ with a SKILL.md (frontmatter name and description, then instructions). Keep skills short — they cost tokens every time they load.

Run the tests:

python3 -m unittest discover tests

Before you use it

• Token counts are estimates. Characters ÷ 4 is within about 20% for English text and code. Other languages can use 2–4× more tokens, especially Chinese, Japanese and Korean. Don’t use these numbers for billing.
• The thresholds are defaults, not your plan’s limits. Prices and limits differ by model and plan. Edit rule 4 in rules/always-on.md to match yours.
• The plugin runs a shell hook. It only prints the rules file. Read hooks/hooks.json before installing, as you should with any plugin.
• Always-on costs about 250 tokens per session. That’s the price of the rules staying in context.
• Review rewritten prompts. Guesses marked [ASSUMED] may be wrong. Never put passwords, keys or private data into prompts.
• Requirements. The estimator needs Python 3. It only reads files locally and sends nothing anywhere. On Windows, the hook runs through Git Bash, which Claude Code uses there. Antigravity folder paths vary between versions.

License

MIT. Provided as-is, without warranty.

Not affiliated with or endorsed by Anthropic or Google. Claude, Antigravity and Gemini are trademarks of their respective owners.