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 promptlint-mcp promptlint <file>

# CLI: lint stdin
echo "be concise and thorough" | npx -p promptlint-mcp promptlint -

# MCP server (auto-launched by Claude Code via mcp config; see below)
npx -y promptlint-mcp

From source (hacking on it)

git clone https://github.com/sean-sunagaku/promptlint-mcp.git
cd promptlint-mcp
npm install
node src/cli.mjs examples/bad-prompt.md

Requires Node >= 18. Zero runtime deps other than the MCP SDK.

CLI usage

Default human report

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] "thanks" fluff (match 1 of 2)
         → Delete. AIs do not need politeness tokens.
         · Thank you
  info   [trailing-fluff] "please" fluff (match 1 of 1)
         → Delete. AIs do not need politeness tokens.
         · Please
  info   [trailing-fluff] "hope-helps" fluff (match 1 of 1)
         → Delete. AIs do not need politeness tokens.
         · I hope this helps
  ...

A clean prompt (see examples/good-prompt.md) returns score 100/100 with zero issues — silence is a feature.

--json for machine consumption

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

{
  "score": 14,
  "issues": [
    {
      "rule": "ambiguous-pronoun",
      "severity": "info",
      "message": "Pronoun \"that\" has no clear referent (line 1)",
      "line": 1,
      "suggestion": "Replace with a concrete noun, e.g. \"this\" → \"the config file\"."
    },
    {
      "rule": "contradiction",
      "severity": "error",
      "message": "Contradicting directives (length): \"concise\" vs \"detailed\"",
      "suggestion": "Pick one, or scope each with a clear condition (e.g. by task type)."
    }
  ],
  "trimmed": "…",
  "tokensBefore": 367,
  "tokensAfter": 307,
  "saved": 60,
  "savedPercent": 16
}

Empty input adds "note": "empty input". Pure-fluff input (e.g. "Thanks!") that would otherwise trim to empty instead returns the original with "trimmerFallback": true — a safe drop-in never hands a blank prompt to your LLM.

--trim to print trimmed output (for piping)

echo "You are a helpful coding assistant. Please always be concise. Thank you for your help! When the user asks a question, handle it carefully. If they provide code, review it. Let me know if anything is unclear. I hope this helps!" \
  | node src/cli.mjs - --trim

You are a helpful coding assistant. Always be concise. When the user asks a question, handle it carefully. If they provide code, review it.

Fenced code blocks are preserved verbatim. Quoted strings ("…") and inline backticks (`…`) are masked before fluff removal so example text survives untouched.

Stdin

cat prompt.md | node src/cli.mjs -

The report header shows promptlint <stdin>.

Other flags

node src/cli.mjs --rules      # list the 5 rule IDs, one per line
node src/cli.mjs --help       # usage

Exit codes

MCP server usage

Register with Claude Code in one command (no clone, no global install needed):

claude mcp add promptlint -- npx -y promptlint-mcp

Or add this to ~/.claude.json (user scope) or project .mcp.json (project scope):

{
  "mcpServers": {
    "promptlint": {
      "command": "npx",
      "args": ["-y", "promptlint-mcp"]
    }
  }
}

If you've globally installed the package, the absolute-path form also works and skips npx's first-run download:

{
  "mcpServers": {
    "promptlint": {
      "command": "promptlint-mcp"
    }
  }
}

Tools exposed

Both tools are pure / idempotent. No network, no filesystem, no state.

Rules

Score formula: 100 - (errors*25 + warns*10 + infos*3), then clamped to <=59 if any error exists and <=79 if 2+ warnings exist. Severity ceilings are a traffic-light: any error = red zone, regardless of count.

Pricing & Roadmap

Free (OSS, MIT)

• CLI (node src/cli.mjs) — all three modes (report / JSON / trim), all 5 rules.
• MCP server (node src/mcp.mjs) — both tools, unlimited use.
• No account, no telemetry, no network.

Pro (planned, $9/mo)

• Team-shared custom rule dictionary (your project's bannned phrases, required headers, forbidden contradictions).
• Per-project lint history with diff view.
• Cross-project token-savings dashboard ("your team saved 1.2M tokens this month").

Pro is not shipped yet. No signup form today — 👍 the early-access issue (#1) to register interest. We DM invites in 👍 order.

Development

Smoke test

npm run lint:self         # lints examples/bad-prompt.md

Expected: exit 1 (2 error contradictions), score in the low teens, all rules firing. If this drifts, a rule regressed.

Baseline

examples/good-prompt.md must always score 100/100 with 0 issues. If a change breaks that, the change is wrong.

Adding a rule

Rules live in src/linter.mjs. Each rule is a plain function:

function ruleFooBar(text) {
  // return an array of { rule, severity, message, samples?, suggestion, line? }
  return [];
}

Add it to the issues array inside lint() and append its ID to the RULES export. If it targets prose (not code), run it against the maskCodeBlocks() copy so fenced examples don't trigger it. If it needs sentence splitting, reuse splitSentences (quote-aware) or splitSentencesPreservingDelims (quote-aware, keeps delimiters for the trimmer).

Contradictions

Extend CONTRADICTION_PAIRS at the top of src/linter.mjs. Each entry is:

{ a: ["keyword-a", ...], b: ["keyword-b", ...], label: "short-label" }

Existing labels: length, tone, frequency, prohibition, commenting, asking.

Community

• 💬 Discussions — questions, ideas, prompt war stories
• 🐛 Issues — bugs, feature requests
  • #1 Pro early access (pinned)
  • #2 Roadmap
  • #3 Known issues / day-2 backlog
• ⭐ Star the repo if it saves you tokens — discovery helps other AI users find it.

How this was built

This package was produced end-to-end by two AI agent teams:

• ai-auto-improve-app — built and polished the product. Customer AI uses the tool and writes feedback → Developer AI patches the code → repeat. Each round's commit (round-001 … round-003) is the developer-AI's response to that round's customer-AI feedback.
• ai-launch-team — published and distributed it. 5 sub-agents handled npm publish, GitHub Release, awesome-mcp PRs, Discussions threads, social copy drafts, and SVG/PNG asset generation.

Combined, this took the product from "empty repo" to "v0.1.2 on npm + 2 awesome-mcp PRs + 2 Discussions seed threads + 4 Issues + full social drafts" with the only manual step being the platform-mandated SMS / OTP authentications (X account creation, npm 2FA setup) — the rest was AI.

License

MIT — see LICENSE.