Skip to content

CLAUDE.md

Latest commit

 

History

History
51 lines (33 loc) · 2.91 KB

CLAUDE.md

File metadata and controls

51 lines (33 loc) · 2.91 KB
Raw

Claude Code — Project Context for AICertify

Read AGENTS.md first. This file inherits all of those instructions and only adds Claude Code-specific notes.

What you're working on

AICertify evaluates AI applications against regulatory frameworks (EU AI Act, NIST AI RMF, +6 more) using OPA policies from gopal. The deliverable is an audit-ready compliance report (PDF / Markdown / JSON / HTML).

Fast orientation

When the user asks a question about the codebase, in order:

  1. For "how does the eval flow work?" → read examples/quickstart.py. It's the canonical user path.
  2. For "what's covered by regulation X?" → look in aicertify/opa_policies/<domain>/<framework>/v1/.
  3. For "what does a report look like?" → open one in examples/outputs/.
  4. For "what's the public Python API?" → read aicertify/__init__.py — it's the surface contract.

Conventions Claude Code should respect

  • Use TodoWrite for multi-step tasks (this is the user's preferred tracking).
  • Run the quickstart before claiming a feature works end-to-end. Type-check and tests verify code correctness; the quickstart verifies user-facing correctness.
  • When editing OPA policies in aicertify/opa_policies/, remind the user that these are vendored from gopal and the upstream copy should usually be edited first.
  • Reports — when changing report templates, open the generated PDF (examples/outputs/eu_ai_act/) and inspect it visually. Don't assume rendering correctness from the code alone.

Useful skills

This project ships Claude Code skills in skills/. To register them, copy into your Claude Code skills directory:

mkdir -p ~/.claude/skills && cp -r skills/* ~/.claude/skills/

Available slash commands once registered:

  • /run-compliance-check — Run the end-to-end quickstart and surface the generated report
  • /evaluate-contract — Evaluate a user-supplied contract JSON against any supported framework
  • /explain-regulation — Walk every policy in a framework directory and explain what it checks
  • /draft-policy — Scaffold a new Rego policy with metadata, default rule, and a test sibling

See skills/README.md for the full list and conventions.

What NOT to do in this repo

  • Don't run python examples/quickstart.py repeatedly without cleaning reports/ — the directory accumulates artifacts.
  • Don't add Python ≥3.13 syntax (project pins to 3.12).
  • Don't switch package manager — poetry is canonical, pip install -e . works for users.
  • Don't claim MCP-compatibility yet — an MCP server is on the roadmap, not shipped.

Session etiquette

The author prefers terse, surgical responses. Present the plan first when ambiguous; ask before introducing new abstractions.