HarnessEngineeringSkills

Harness Engineering Skills

Skills that encode the practices of harness engineering — the discipline of designing environments, constraints, and feedback loops so AI coding agents can work autonomously at speed.

The core idea: stop writing code, start building systems. When your attention is the only scarce resource, the leverage comes from encoding standards once and letting agents execute.

Background

Harness engineering emerged from OpenAI's Frontier team, where three engineers built a production app with zero human-written code over five months. The key insight wasn't about better prompts — it was about better scaffolding: specs, fast build loops, observability, and automated review.

• Harness Engineering — OpenAI's original post on the discipline
• Latent Space interview — Deep dive with Ryan Lopopolo on the concrete practices
• Go Faster — The motive: speed through systems, not effort

Skills

Skill	Description
setup-harness	Scaffold a new repo for agent-first development — docs, harness scripts, CI, smoke tests, and automerge workflows
update-harness	Audit and improve an existing agent harness without replacing local conventions

Install

npx skills add Swiftyos/HarnessEngineeringSkills@setup-harness
npx skills add Swiftyos/HarnessEngineeringSkills@update-harness

Principles

These skills encode a few hard-won ideas:

1. Agents run scripts, not ad-hoc commands. Scripts are the source of truth. If an agent keeps making the same mistake, the fix is a better script — not a better prompt.

2. One-minute build loops. Fast feedback keeps agents productive. When builds get slow, refactor the architecture rather than relaxing the constraint.

3. Docs fail CI. Generated docs have freshness checks. Links are validated. Every committed directory carries an INDEX.md contract that must stay current.

4. Start conservative on automerge. Green CI + agent review + human merge. Widen only after the repo proves stable.

5. Code is disposable, structure is not. Agents can regenerate code from specs. What matters is the scaffolding: the docs, the checks, the feedback loops.

License

MIT