The keyword trigger is the interesting part — it keeps the tool from firing on every message and makes refinement something you ask for explicitly. The scope is genuinely one job, so there is nothing to learn beyond the two parameters.
A single-purpose MCP server that takes a draft prompt and returns an engineered version of it — structure, explicit requirements, domain context — using Anthropic's Claude behind the scenes. It is a writing step you can call from inside any MCP client rather than a separate prompt-engineering app.
- `rewrite_prompt` — hand it a `prompt` string and get back a restructured version with clearer instructions, stated expectations and formatting
- A `language` parameter for code-related prompts, so the rewrite carries programming-language-specific considerations
- A deliberate trigger: the input must contain the keyword "refine", so the rewrite only fires when you mean it
An Anthropic API key, supplied as `ANTHROPIC_API_KEY`; without it the tool returns an error. The package.json names it `refine-prompt` at 1.2.0, which npm does not have; a clone plus `npm install` and `node index.js` works. It speaks stdio.