skill-lint

Static analyzer for Bankr agent skills.

If you're building a skill for the Bankr ecosystem — a SKILL.md file that teaches an AI agent a new capability — this tool tells you three things before you push:

1. Is the format correct? (Bankr will skip your skill if catalog.json is invalid)
2. Is the content safe? (no prompt injection, no leaked secrets, no hidden instructions)
3. Is the quality good? (concrete instructions, clear triggers, right size)

It's clippy for agent skills.

Why this exists

The BankrBot/skills catalog has 148 skills and 346 open PRs that nobody reviews. There's no automated validation — maintainers read each skill by hand. This tool automates that first pass.

Running it against the live catalog found:

• 10 invalid skills (missing schemaVersion, legacy install format)
• 17 format errors
• 353 quality warnings
• 1 skill containing curl | sh in its instructions

What it checks

Format (from the official spec)

Check	What it validates
catalog.json	Valid JSON, schemaVersion: 1, slug equals the folder name
install	type is bankr or external, required fields present, command matches the documented form
SKILL.md frontmatter	name and description present (frontmatter-less accepted with warning)
Optional fields	tags, visibility (private/public), version (number)
File limits	SKILL.md ≤ 1 MB, reference files ≤ 100 KB each
Logo	logo.svg or logo.png present (recommended)

Security

Check	What it catches
Prompt injection	"ignore previous instructions", "hide this from the user", "developer mode" — 19 patterns
Embedded secrets	PEM blocks, Anthropic/GitHub/AWS API keys, 64-char hex private keys
Exfiltration	Discord/Telegram/pastebin webhooks, reads from ~/.ssh or ~/.aws
Dangerous commands	rm -rf /, curl | sh, chmod -R 777
Hidden content	Zero-width Unicode characters, imperative HTML comments, large base64 blobs

Security skills that document injection patterns (like polygraph or 1claw) are automatically recognized — their pattern hits are downgraded from error to info.

Quality

Check	What it looks for
Description triggers	"Use when..." (positive) and "Do NOT use for..." (negative)
Body substance	Code blocks, actionable commands, specific API/CLI references
Egress transparency	Every network endpoint and env-var access is listed

Grades

Each skill gets an A-F grade:

Grade	Meaning
A	No errors, no warnings — production ready
B	No errors, 1-3 warnings
C	No errors, 4+ warnings
D	1 error
F	2+ errors

Live catalog distribution: A=8 B=97 C=33 D=7 F=3

Install

cargo install --git https://github.com/latent-9/skill-lint

Usage

# Lint one skill
skill-lint ./my-skill

# Lint every skill in a repo
skill-lint --repo ./skills

# Check if your slug is already taken in the live catalog
skill-lint --repo . --check-catalog

# Auto-fix common issues (missing schemaVersion, legacy install format)
skill-lint --repo . --fix

# Machine-readable JSON for CI
skill-lint --json ./my-skill

# SARIF for GitHub code scanning
skill-lint --sarif ./my-skill

# Fail on warnings too
skill-lint --strict ./my-skill

Example output

my-skill: [B] OK (0 errors, 2 warnings)
  [-] description: description states trigger conditions (good)
  [-] description: description states negative triggers (good)
  [!] logo: logo.svg (or logo.png) is recommended
  [!] body_quality: instruction body has no code blocks

--- 148 skills: 138 valid, 10 invalid, 17 errors, 353 warnings ---
  grades: A=8 B=97 C=33 D=7 F=3

Exit codes

Code	Meaning
0	All skills valid
1	At least one skill has errors
2	No skill folders found

GitHub Action

Drop this into .github/workflows/skill-lint.yml:

name: skill-lint
on: [pull_request]
jobs:
  lint:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - run: cargo install --git https://github.com/latent-9/skill-lint
      - run: skill-lint --repo .

Or use the template in this repo.

Library

The validation logic is a library crate — embed it in your own tool:

use skill_lint::lint_skill;

let report = lint_skill(std::path::Path::new("./my-skill"));
println!("grade: {}", report.grade.as_str());
println!("valid: {}", report.is_valid());

Bankr ecosystem

• Bankr — AI agents that fund themselves via token trading fees
• BankrBot/skills — the public skill catalog (148 skills, 346 open PRs)
• SKILL.md format reference — the official spec this tool validates against

License

MIT