promptlint-mcp

Lint AI prompts like code. Save tokens. Catch contradictions.

A static analyzer for AI prompts — system prompts, agent instructions, tool descriptions — that runs as a CLI and as an MCP server so Claude Code (or any MCP-aware agent) can lint prompts before sending them.

Quick demo

Run promptlint on a typical (silently broken) system prompt:

$ node src/cli.mjs examples/bad-prompt.md

promptlint examples/bad-prompt.md
  score  14/100
  tokens 367 → 307 (-60, 16% saved)
  issues 14

  info   [ambiguous-pronoun] Pronoun "that" has no clear referent (line 1)
         → Replace with a concrete noun, e.g. "this" → "the config file".
  ...
  error  [contradiction] Contradicting directives (length): "concise" vs "detailed"
         → Pick one, or scope each with a clear condition (e.g. by task type).
  error  [contradiction] Contradicting directives (commenting): "do not add comments" vs "explain each step"
         → Pick one, or scope each with a clear condition (e.g. by task type).
  info   [trailing-fluff] "please" fluff (match 1 of 1)
         → Delete. AIs do not need politeness tokens.
         · Please
  ...

Two error-level contradictions caught, score clamped to 14/100, 60 tokens (16%) reclaimable via --trim.

And here's --trim in action on a small input:

$ echo "You are a helpful assistant. Please always be concise. Thank you for your help! When the user asks a question, handle it carefully." | node src/cli.mjs - --trim

You are a helpful assistant. Always be concise. When the user asks a question, handle it carefully.

See docs/promo/ for posting drafts and the (TODO) asciinema cast.

Why

• Contradictions: "be concise" + "be thorough" ships silently and wastes tokens on every request. promptlint flags paired opposites (length / tone / frequency / commenting / asking) as error.
• Redundancy: near-duplicate sentences (Jaccard > 0.6) get one warn apiece — delete one.
• Ambiguous pronouns: "use it" / "handle that" without a mid-sentence referent is flagged with a line number. Start-of-sentence capitals do NOT count as referents, so normal English doesn't false-positive.
• Long code examples: fenced code blocks over ~1200 chars or 30 lines trigger a warn — examples are almost always trimmable to shape-only.
• Politeness fluff: Please / Thanks / I hope this helps / Let me know if… / Feel free to… are detected, counted, and (with --trim) stripped while keeping your actual imperatives intact.

Scope: English prompts, prose — not source code. A Japanese prompt or a file of JavaScript will look clean even if it isn't.

Install

Install both the CLI and the MCP server in one go:

npm install -g promptlint-mcp

Two binaries are placed on your $PATH:

Run without installing

# CLI: lint a file
npx -p pro

... [View full README on GitHub](https://github.com/sean-sunagaku/promptlint-mcp#readme)