Audit Ur Harness

Audit ur Harness

An [Agent Skill](https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills) that audits a project's AI **harness** — the guardrails, memory, verification, observability, and tool-design scaffolding around a model, not the model itself — and produces an evidence-cited maturity report. It never edits your project and never applies its own suggestions; it only ever reads files and writes a report.

What it does

Point it at a project folder and ask it to audit the harness. It reads through the project, scores each of 8 fixed layers (State & Memory, Execution Environment, Control & Enforcement, Context Management, Reasoning & Verification, Tool Design, Long-Horizon Execution, Observability & Learning) on a 0–3 scale, and writes a self-contained interactive HTML report — every score backed by a citation to the actual file and line it came from. "No evidence found" is a valid, complete answer for any layer; nothing is invented.

It can also notice a real, evidenced pattern specific to your project that doesn't fit any of the 8 fixed layers, score it the same way, and suggest — never apply — adding it as a permanent custom layer for future audits.

Why

Most "AI harness" advice is generic. This turns it into a repeatable, grounded audit: concrete citations instead of vibes, absence treated as a real finding instead of being glossed over, and every improvement suggestion left as a copy-pasteable draft for a human to review and apply through their own agent — never auto-applied, under any setting.

Install

This is a standard [Agent Skill](https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills) — a `SKILL.md` file plus a `references/` folder. It works in any tool that loads Agent Skills (Claude Code, Cowork, Codex, Cursor, and others).

1. Download this repository (or the latest release).
2. Copy the Audit_ur_Harness/ folder into wherever your tool looks fo