Philosophy

Long-running agent work fails in predictable ways:

1. Context drifts. The agent forgets the goal halfway through.
2. Verification is skipped. "It compiles" becomes "it works."
3. Safety rules live in chat history. Nothing enforces them.
4. Skills are re-explained every session. The same prompt boilerplate is rewritten.
5. Reviews happen too late. Mistakes propagate.

claude-harnesses is a deliberate response to each:

1. Ledger captures goal, plan, progress, decisions, risks, verification — durable across sessions and compactions.
2. Stop-verify hook blocks completion until scripts/verify.sh passes. PostToolUse hooks run format/typecheck/test on every edit.
3. Permissions presets + safety-pack PreToolUse hooks enforce safety mechanically.
4. Skills with frontmatter give Claude Code reusable workflow definitions; install once.
5. Review-pack + pr-guardian-pack + tdd-pack put review and test discipline in front of merge, not after.

Harnesses do not replace judgment. They keep the cheap mistakes from happening so judgment can focus on the expensive ones.