harness-doctor

KO | EN

gitlite — search

#cross-platform #rag #codex #css #awesome #devops #ui #api #mysql #cli

(★ 9)

A Claude Code skill that audits your agent harness — contract, skills, guides/sensors, hooks, memory — and reports prioritized fixes.

File Explorer

ZIP Download

- checklist.md
- principles.md
- scan.sh
.gitignore
LICENSE
README.md
SKILL.md

# Use via CDN

jsDelivr

jsDelivr serves any public GitHub repository as a CDN with zero setup. Pick a version and a file to get a ready-to-paste link and snippet.

Version

File

Link

Example

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

A Claude Code **skill that audits your agent harness** — and tells you what's good, what's missing, and what to fix first.

> **Agent = Model + Harness.** The model is one box. *Everything else* — your `CLAUDE.md` contract, your skills, your tests and linters, your hooks — is the harness, and it's the part you build. Harness Doctor grades it.
> Framing after Birgitta Böckeler, [*Harness Engineering*](https://martinfowler.com/articles/harness-engineering.html) (martinfowler.com).

It's also a worked example of a **robust skill**: a lean `SKILL.md` orchestrator that loads `references/checklist.md` on demand (progressive disclosure). It audits the very thing it's made of.

## What it checks

| Dimension | It looks for |
|---|---|
| **Contract** | `CLAUDE.md` exists, has hard *always/never* rules, is **< ~200 lines**, holds facts not how-to |
| **Skills** | descriptions written **WHEN-not-WHAT**, gerund names, progressive disclosure, no overlaps |
| **Guides** *(feedforward)* | conventions, specs, architecture docs that steer *before* the agent acts |
| **Sensors** *(feedback)* | tests, linter, type-checker, CI that check *after* — and the feedforward/feedback balance |
| **Enforcement** | `PreToolUse` **hooks** for hard guarantees (a rule is only hard if a hook blocks it) |
| **Memory & State** | durable decisions/progress that survive context resets; long-task state |
| **Improvement loop** | the harness *gets better* — evals, fail→fix→encode, drift monitoring |

Output is a scannable **scorecard + prioritized fixes** (cheapest, most deterministic first — *start cheap*), then it offers to apply the easy ones.

## Install

**One command — any agent** (Claude Code, Codex, Cursor, Gemini CLI, OpenCode…):
```bash
npx skills add jubscodes/harness-doctor
```

**Or drop-in for Claude Code:**
```bash
git clone https://github.com/jubscodes/harness-doctor.git \
  ~/.claude/skills/harness-doctor && rm -rf ~/.claude/skills/harness-doctor/.git
```

## Use

In any repo:
```
/harness-doctor
```
or just ask: *"audit this repo's harness."* It scans, scores, reports, and offers fixes.

## How it's built

```
harness-doctor/
├── SKILL.md              # the orchestrator — process + "load X when Y"
└── references/
    ├── checklist.md      # the audit criteria, loaded on demand
    └── principles.md     # the why + sources (Böckeler · ECC · Anthropic · skills.sh)
```

That's the anatomy of a strong skill: the orchestrator stays lean; the depth lives one hop away.

## Credits

- **Harness Engineering** framing — Birgitta Böckeler / Thoughtworks · https://martinfowler.com/articles/harness-engineering.html
- **When to use what / skill best practices** — Claude Code docs · https://code.claude.com/docs
- Built by [Jubs](https://jubs.studio) — `get-shit-pretty` author, product engineer.

## License

harness-doctor

Owner Repos

File Explorer

# Use via CDN

Was this content helpful?

Recent Feedback

harness-doctor

Owner Repos

File Explorer

# Use via CDN

Was this content helpful?

Recent Feedback

Download README