Skip to content
Anvil

Agent Skill

Copy the skill below into your agent’s instruction file — CLAUDE.md, .cursor/rules/anvil.mdc, AGENTS.md, or any context file your agent reads.

---
name: anvil
description: Audit AI rule files for coverage gaps, drift, conflicts, and format compliance. Use Anvil when you need to evaluate or improve the quality of AI instructions in a codebase.
---


# Anvil — AI Rules Audit


## When to use


- Starting work in a new repo — audit the existing AI rules to understand what's covered and what's missing
- After modifying AI rule files — verify the changes didn't introduce drift or conflicts
- Periodically — catch stale globs, broken references, and coverage gaps before they cause problems
- When onboarding — bootstrap starter rules for a repo that has none


## Install


Zero-install. The default audit path expects either a local AI CLI login or an `OPENAI_API_KEY`; use `--ci` when you specifically want the local-only path.


## Quick audit


`bunx @lambdacurry/anvil audit --target .`


This runs the full product path: discovers rule files, detects drift, scores coverage, and adds AI-synthesized improvement priorities to the markdown report.


## Save the report


`bunx @lambdacurry/anvil audit --target . --output ./audit-report.md`


## CI / local-only mode


If you want a deterministic local-only structural lint pass:


`bunx @lambdacurry/anvil audit --target . --ci`


Auto-detects available local AI CLIs (Claude Code, Codex, Gemini CLI, opencode) or OpenAI by default when you omit `--ci`.


Force a specific provider: `--ai-provider claude-code|codex-cli|gemini-cli|opencode|openai|heuristic`


## Other commands


- `bunx @lambdacurry/anvil drift .` — detect drift only
- `bunx @lambdacurry/anvil bootstrap . --output ./bootstrap-draft.md` — generate starter rules
- `bunx @lambdacurry/anvil mine-pr owner/repo` — mine PR history for missing rules (requires gh CLI)


## What to look for in the report


- **Rule Quality Score (0–100)** — overall health of AI rules on the full AI-backed path
- **Structural Lint Score (0–100)** — local-only score label when you run with `--ci`
- **Guardrail Readiness Score (0–35)** — engineering guardrail coverage
- **Coverage gaps** — patterns with no rule protection
- **Drift issues** — stale globs, broken paths, outdated references
- **Recommendations** — prioritized improvements


## Trust model


- `--ci` = fully local, zero outbound calls, no data leaves the machine
- Default `audit` path = auto-detects local AI CLIs first, then OpenAI if API key is set
- `--no-ai` = hidden deprecated alias for `--ci`
- Anvil never modifies files in the target repo (read-only audit)


## Repo config


Create `.anvil/config.yml` in the target repo to tune scoring:


```yaml
version: 1
profile: internal-tool  # or: library, production-app, prototype
hardGates:
  dimensions:
    ciDiscipline:
      minScore: 3
```

Full docs for agents

Section titled “Full docs for agents”

If you want an agent to have access to the complete Anvil documentation (all guides, reference, and configuration), fetch the combined markdown bundle:

Terminal window
curl -s https://lambda-curry.github.io/anvil/llms-full.txt

For a concise index with links to individual pages:

Terminal window
curl -s https://lambda-curry.github.io/anvil/llms.txt

Both follow the llms.txt convention — a standard way for AI agents to discover and consume documentation.