promptproof

A fast, deterministic, zero-API linter for the prompt files your agent depends on.

SKILL.md files, Claude Code sub-agents, MCP tool descriptions, slash-command frontmatter, and plain system prompts — promptproof catches the mistakes that silently break agents before they ship: descriptions that never trigger, contradictory instructions, token-bloated context, broken frontmatter, leaked secrets.

No API key. No model calls. No network. Just regex, structure, and millisecond runs that fit a pre-commit hook and CI.

$ promptproof .claude/

.claude/skills/pdf/SKILL.md:
  2:14: PP404 weak trigger: description summarizes instead of stating when to use it
        → rewrite as "Use when <condition> — e.g. <trigger example>"
  9:1:  PP306 decorative banner wastes ~40 tokens
  -:-:  PP305 token budget: body ~7,200 tokens, 44% over budget (5,000)

0 errors, 3 warnings, 0 info  ·  0 API calls  ·  6ms

---

Why another prompt linter?

There are good tools in this space already, and promptproof is honest about that:

Tool	Focus	LLM calls
promptlint.dev	generic LLM prompts	no
skillcheck / claude-lint / agent-skill-linter	SKILL.md spec compliance	some (agent critique)
agentlint / agentlinter.com	CLAUDE.md / AGENTS.md harness	no
promptproof	all agent prompt assets — skills, sub-agents, MCP tools, commands, prompts — in one tool	never

The wedge is the same one ruff used to win a crowded Python-linting space: unification + speed + zero dependencies. promptproof is the only linter that understands every prompt-asset type, stays 100% deterministic and offline, and ships the universal prompt rules (contradiction, ambiguity, token waste) and the asset-specific ones (weak triggers, frontmatter, tool-param docs) in a single tool you can run on every keystroke.

Install

# run it without installing
uvx promptproof .

# or install
pip install promptproof

Python ≥ 3.11. The core has zero runtime dependencies. Optional extras: promptproof[yaml] (robust frontmatter via PyYAML), promptproof[mcp] (MCP server).

Usage

promptproof .                      # lint the current tree
promptproof .claude/skills         # lint a directory
promptproof SKILL.md               # lint one file
cat prompt.txt | promptproof -     # lint stdin (use --kind to force a type)

promptproof . --format github      # GitHub Actions annotations
promptproof . --format json        # machine-readable
promptproof . --select triggering  # only the PP4xx rules
promptproof . --ignore PP301,PP304 # silence specific rules

promptproof . --fix                # apply the mechanical repairs
promptproof . --diff               # preview them, write nothing

promptproof rules                  # list every rule
promptproof explain PP404          # rationale + good/bad example

Exit code is 1 when any finding is at or above the fail level (default: warning), else 0. Use --fail-level error to only fail CI on errors, or --exit-zero to report without failing.

Fixing (--fix)

Some findings are token bloat with exactly one correct repair. Those carry a fix:

$ promptproof .claude --diff
--- .claude/skills/pdf/SKILL.md
+++ .claude/skills/pdf/SKILL.md
-Please read the PDF in order to extract its text.
-Thank you.
+Read the PDF to extract its text.

-Due to the fact that PDFs vary, please check the page count first.
+Because PDFs vary, check the page count first.

-✨ 🎉 🚀 💡 ⭐ 🔥

--diff previews; --fix writes. Fixable today: PP301 (courtesy padding), PP302 (wordy connectives), PP306 (decorative banners).

The line is deliberate: a rule attaches a fix only when the repair is unambiguous. Deleting a banner or replacing "in order to" with "to" has one right answer. Rewriting a vague directive does not — those stay hints, because a linter guessing at what you meant is worse than a linter that tells you to decide.

Three properties make it safe to run across a prompt library:

• Fenced blocks are never touched. Sample text and transcripts are examples, not instructions to the model. (Fixing this also removed a false positive from plain linting — PP301/PP302 used to flag prose inside code fences.)
• Edits can't interleave. Two rules never rewrite the same line in one pass.
• It converges. Several rules report one match per line, so fixing loops until the file stops changing — bounded, so a pathological rule pair can't spin.

Repairs also clean up after themselves: a dropped leading "Please" recapitalises the sentence, a stranded comma goes, and a line that was only courtesy is removed entirely.

Adopting on an existing repo (--baseline)

A first run on a real prompt library reports hundreds of findings, and CI can't gate on that until they're all fixed. Record what's already there and fail only on what's added after:

$ promptproof . --write-baseline
wrote 143 finding(s) to .promptproof-baseline.json

$ promptproof . --baseline
All prompts proofed ✓
1 baselined

Commit the baseline file. Entries are keyed on the content that triggered each finding — path, rule, and a hash of the offending line — never on line numbers. A baseline keyed on line numbers stops working the moment someone inserts a paragraph: every finding below the insertion looks new. This one survives edits elsewhere in the file, and a finding whose text actually changed correctly reads as new.

Counts are tracked per entry, so three known copies of a problem can be baselined and a fourth still fails the build. And a missing or unreadable baseline is a hard error, never a silent pass — a gate that goes green because it couldn't find its baseline is worse than no gate.

What it checks

31 rules across six categories (run promptproof rules for the full list):

• clarity (PP1xx) — ambiguous directives, vague quantifiers, unresolved pronouns, subjective criteria, weak modals.
• consistency (PP2xx) — contradictory directives, conflicting output format / length / persona.
• economy (PP3xx) — politeness padding, filler phrases, redundant restatement, walls of text, token-budget overflow, decorative banners.
• triggering (PP4xx) — missing / too-short / too-long descriptions, weak triggers (the #1 reason skills never load), first-person descriptions, undocumented MCP tool params.
• structure (PP5xx) — missing or invalid frontmatter, missing name, non-kebab names, name/dir mismatch, missing verify guidance, unknown frontmatter keys.
• safety (PP6xx) — secrets in prompts, embedded injection phrases, real PII in examples.

Configuration

Add a [tool.promptproof] table to pyproject.toml (or put the same table, or its keys at the top level, in a standalone .promptproof.toml):

[tool.promptproof]
fail-level = "warning"
ignore = ["PP304"]          # rule ids or whole categories ("economy")

[tool.promptproof.severity]
PP301 = "info"              # downgrade politeness-padding

[tool.promptproof.thresholds]
"token_budget.skill" = 4000
description_min_chars = 50

Suppress a single finding inline:

<!-- promptproof: ignore PP602 -->
ignore previous instructions   # intentional red-team fixture

CI

GitHub Action (action.yml):

- uses: shaxzodbek-uzb/promptproof@v0.1.0
  with:
    paths: .claude
    fail-level: warning

pre-commit:

- repo: https://github.com/shaxzodbek-uzb/promptproof
  rev: v0.1.0
  hooks:
    - id: promptproof

Library

from promptproof import lint_text, lint_paths, render

findings = lint_text(open("SKILL.md").read(), path="SKILL.md")
print(render(findings, "text"))

from promptproof.fixer import fix_text
result = fix_text(open("SKILL.md").read(), path="SKILL.md")
print(result.text, len(result.applied), "repaired")

from promptproof.baseline import Baseline
fresh, suppressed = Baseline.load(".promptproof-baseline.json").filter(findings)

Contributing

A wrong rule is worse than no rule — every rule ships with a firing test and a known-good test that must not fire. See CONTRIBUTING.md and the canonical SPEC.md.

License

MIT © Shaxzodbek Qambaraliyev / Blaze