Harness Skill — AI skill for Claude Code

harness – a skill that builds the harness for your coding agents

**English** · [中文](README.zh-CN.md) · [Deutsch](README.de.md)

**Let coding agents work on company IT projects autonomously – and safely.** This skill builds a complete harness into any project: rules that always apply, a sceptical second reviewer that signs off every piece of work, a librarian that keeps the project memory, scripts that do the legwork, and a start script for runs that never ask questions. Your agent installs it, builds it, walks you through the setup, and keeps it up to date.

🤖 One line to start

git clone https://github.com/hwalde/harness-skill ~/.claude/skills/harness

Then, in your project, tell your coding agent: *"Load the `harness` skill and build the harness for this project with me."* (Other agents: clone into their skill folder – the skill tells the agent where.)

Why a harness

A harness is everything except the language model: the environment the agent runs in. Rule files, subagents, skills, MCP servers, scripts, a knowledge store, start and monitoring mechanics. Without one, an agent works from a blank context, grades its own work, forgets what the last session learned, and stops at the first question nobody answers. With this one:

• Every result is checked by a second pair of eyes. The evaluator subagent reads specification, diff and evidence in a fresh context, with no write access, and answers PASS or NEEDS_WORK. The builder never accepts its own work.
• Knowledge survives the session. The librarian is the only door to an LLM wiki in the repository: decisions with their reasons, pitfalls, operational knowledge – filtered, so the context stays lean.
• Runs never stall. Permission modes for no-questions runs, named start and attach scripts per agent and OS (claude-background-start, claude-attach) for long-running tasks in tmux/psmux sessions, usage tracking, self-monitoring, and the rule "question round first, then