damson

Agent Config Harness — AI skill for Claude Code

AI community

Agent instruction files, scored and regression-tested like code.

How to install Agent Config Harness

This entry records only its repository, not the path inside it, so there is no exact command to give. Open damson/agent-config-harness and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Agent Config Harness does

Agent instruction files, scored and regression-tested like code. CLAUDE.md / AGENTS.md / skills managed via a domain registry, an LLM rubric, and a bats suite.

Alternatives in AI

Part 4 — Generate AGENTS.md And AI Agent Configuration Files — I'll help you create the instruction files that will guide your AI coding assistant to build your MVP 2.1k ★
Ante — Ghost in your shell 1.9k ★
Mancode — AI coding agent harness 358 ★

README

agent-config-harness

**Agent instruction files, scored and regression-tested like code.**

`CLAUDE.md`, `AGENTS.md`, `.cursorrules` and skill definitions steer how coding agents behave, and they are usually maintained the way prompts are maintained: edited by feel, never measured, quietly rotting. This repo treats them as something you can grade, track, and break a build over.

$ just eval backend-node
ℹ Scoring domain: backend-node
✔ Result written: evals/results/2026-08-29-backend-node.json
ℹ   Total: 24/25  Grade: A
ℹ   Findings (5):
    - [completeness] No build/run/verify commands. An agent cannot tell how to
      install, typecheck, lint or run the test suite before claiming a change works.
    - [clarity] Logging: `requestId` propagation is required but the mechanism is
      unspecified — AsyncLocalStorage vs an explicit context argument produces
      very different code.
    …

That is a real run against the example domain in this repo, not an illustration.

Three things make that possible, and they are the reason this exists:


A rubric with a number	Config files are scored 1–5 on clarity, conciseness, completeness, consistency and actionability. Out of 25, with a letter grade and specific findings — not a vibe.
A trend, not a snapshot	Scores are recorded per domain over time, so you can tell an improvement from a rewrite that felt productive.
Structural tests	A `bats` suite enforces the invariants a rubric cannot see: skill frontmatter matching its folder, required headings, unique names across a flat install namespace, no symlink silently turned back into a duplicate file.

**Requires an API key.** The eval harness shells out to the Claude CLI. Everything else — linking, health checks, the test suite — runs offline. If you only want the structure and the tests, you never need a key. And seeing the output needs nothing at all: [evals/examples/](evals/examples/) holds a real A and a real F for the s