kidowkidow/skills/setup-harness/SKILL.md

setup-harness

Skill2 starsChanged 6 months agoMITView on GitHub

Build agent harness infrastructure (Constrain, Inform, Verify, Correct) on a project step by step. Generates AGENTS.md as the canonical rules file with a CLAUDE.md import stub, configures Claude Code permissions and a session-end verify hook, scaffolds optional CI gates, and seeds a Past Failures section for the "1 failure = 1 prevention rule" loop. Optimized for Claude Code; AGENTS.md is the cross-tool standard so OpenAI Codex reads the same file. Use when user wants to set up agent guardrails, harden an AI coding agent setup, write or improve AGENTS.md/CLAUDE.md, configure agent permissions/hooks, or asks "how do I make my agent safer", "set up harness engineering", "configure agent rules for this project".

Reads credentials
Deletes or force-pushes
Commits and pushes

2 other projects use a version of this.

Adopted by

Most recently taken up by

Firzus/agent-skillstheir own version, first seen Oct 13, 2026
LeadGrowGTM/loop-engineertheir own version, first seen Oct 8, 2026

Other versions2 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 2
LeadGrowGTM/loop-engineer★ 13same name
Firzus/agent-skills★ 4same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository17 other files it gives its agents.
- create-nextjs-specificationSkill
- create-product-demoSkill
- find-github-repoSkill
- find-npm-packageSkill
- image-styleSkill
- nextjs-route-scaffoldSkill
- perfect-promptSkill
- place-meSkill
- review-meSkill
- setup-aeo-geoSkill
- songwritingSkill
- step-by-stepSkill
- stock-analysisSkill
- study-meSkill
- teach-meSkill
- three-linesSkill
- visualstormingSkill

The file

We split this skill into the rules it gives an agent

11 rules. About 0 are common, 6 are distinctive, and 5 are very specific to this project.

≈0
Common · estimate
Many files like this have a rule of this kind.
≈6
Distinctive · estimate
This author's own choices. Worth a look.
5
Very specific
Tied to their project's names, files or tools. Probably skip.

Common
Distinctive
Very specific

Very specific:
Write CLAUDE.md containing one line: @AGENTS.md (Claude Code import directive).
line 15
Very specific:
Create .claude/settings.json from templates/settings.json.
line 23
Very specific:
strict: Stop hook runs pnpm lint && pnpm typecheck && pnpm test --run.
line 33
Very specific:
Never use PostToolUse to run tests/typecheck — runs on every edit, kills throughput.
line 35
Distinctive:
After any agent mistake, ask user: "Add a one-line prevention rule to Past Failures?"
line 43
Distinctive:
No edit-count hook, no automated doom-loop counter — agents do not self-track edits, and a hook adds maintenance with little real-world …
line 46
Distinctive:
Failure → rule: every confirmed agent mistake earns one Past Failures line.
line 51
Very specific:
Read before write: never overwrite AGENTS.md, CLAUDE.md, .claude/settings.json, or workflows.
line 52
Distinctive:
Stage skip warning: if user wants Stage 3 before Stage 2, confirm: "Skipping Constrain leaves the agent unbounded.
line 53
Distinctive:
Run Stage 1 → confirm AGENTS.md + CLAUDE.md.
line 59
Distinctive:
Ask: "Continue to Stage 2 (permissions)?"
line 60

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.

Files

SKILL.md

Harness Engineering Setup

Build harness in 3 stages, ordered by impact. Stop after any stage if user is satisfied — each delivers value standalone. See REFERENCE.md for theory.

Stage 1 — Inform (AGENTS.md + CLAUDE.md stub)

Detect project stack from package.json, Cargo.toml, pyproject.toml, go.mod.
Draft AGENTS.md (≤ 60 lines) using templates/AGENTS.md. Pre-fill Project Stack only.
Mark Forbidden and Past Failures as # DRAFT — review required.
Write CLAUDE.md containing one line: @AGENTS.md (Claude Code import directive).
Show diff. Require explicit confirmation before save.
- Warn: "ETH Zurich research: LLM-generated AGENTS.md drops agent performance 3% and raises cost 20%. Past Failures must come from real incidents — leave empty if none."

Pnpm is the standard. AGENTS.md fixes Package manager: pnpm. Refuse to set otherwise.

Stage 2 — Constrain (permissions)

Create .claude/settings.json from templates/settings.json.
permissions.allow: read/edit/write/test/git basics.
permissions.deny: rm -rf *, curl *, wget *, sudo *, chmod 777 *, git push --force *, Read(.env*), Read(**/secrets/**).
No PreToolUse hooks. permissions.deny is the single defense layer for dangerous commands.

Stage 3 — Verify (Stop hook required, CI optional)

Pick one variant from templates/settings.json:

lite (default): Stop hook runs pnpm lint once at session end.
strict: Stop hook runs pnpm lint && pnpm typecheck && pnpm test --run.

Never use PostToolUse to run tests/typecheck — runs on every edit, kills throughput.

Optional: Scaffold .github/workflows/agent-pr-checks.yml from templates/agent-pr-checks.yml. Hooks give immediate feedback per session. CI is the merge gate.

Correct (no extra files)

Failure handling lives inside AGENTS.md Past Failures and the workflow rule:

After any agent mistake, ask user: "Add a one-line prevention rule to Past Failures?"
On repeated failure, restart in a clean context (new session). Do not patch forward.

No edit-count hook, no automated doom-loop counter — agents do not self-track edits, and a hook adds maintenance with little real-world payoff. Add later only after observing repeated incidents.

Operating Rules

60-line rule: AGENTS.md ≤ 60 lines. To add a line, remove a line.
Failure → rule: every confirmed agent mistake earns one Past Failures line.
Read before write: never overwrite AGENTS.md, CLAUDE.md, .claude/settings.json, or workflows. Show diff, ask, then write.
Stage skip warning: if user wants Stage 3 before Stage 2, confirm: "Skipping Constrain leaves the agent unbounded. Proceed?"

Quick Start

User says "set up harness on this project":

Run Stage 1 → confirm AGENTS.md + CLAUDE.md.
Ask: "Continue to Stage 2 (permissions)?"
Continue stage by stage until user stops.

Reference

4 pillars deep dive: REFERENCE.md
AGENTS.md template: templates/AGENTS.md
CLAUDE.md import stub: templates/CLAUDE.md
Settings template (lite + strict): templates/settings.json
CI workflow template: templates/agent-pr-checks.yml

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

It worked
Partly
It didn't

What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

Check your setup, free