First release of llmauditor: a read-only audit of how a codebase integrates Large Language Models. It writes a scored, prioritized, self-contained llmaudit.md at the repo root and prints the verdict in chat. It never calls a model, runs the app, or mutates data or an index.
Dual-compatible: the same SKILL.md runs in both tools.
- Claude Code:
/llmauditor - Codex:
$llmauditor
What it scores
Twelve dimensions (two conditional on the project's surface), weighted and re-normalized so the score is meaningful for a single-call summarizer and a tool-using agent alike:
Security and Trust Boundaries (17), Prompt Construction (12), Model Selection and Routing (10), Provider API and SDK Usage (10), Reliability (10), Output Handling (8), Cost and Quotas (6), Evaluation (5), Observability (5), Latency (4), plus the conditional Retrieval/RAG (7) and Agent/Tool (6) lenses.
Grounded in the OWASP Top 10 for LLM Applications 2025, the OWASP Top 10 for Agentic Applications, NIST AI 600-1, MITRE ATLAS, and provider documentation. A single Critical finding caps its dimension and the overall grade; an ownership map keeps each defect scored once; and the method actively hunts paper controls.
Install
Install the single SKILL.md into whichever tool you use:
```sh
Claude Code
mkdir -p ~/.claude/skills/llmauditor && cp skills/llmauditor/SKILL.md ~/.claude/skills/llmauditor/SKILL.md
Codex (also reads ~/.agents/skills and repo-local .agents/skills)
mkdir -p ~/.codex/skills/llmauditor && cp skills/llmauditor/SKILL.md ~/.codex/skills/llmauditor/SKILL.md
```
Or install the whole repo as a Claude Code plugin (it ships a .claude-plugin/plugin.json manifest). The attached llmauditor-0.1.0.tar.gz contains the full repo.
Generated with Claude Code