Looping Tasks

Install the infrastructure to run Claude Code in an autonomous implementation loop. Each iteration starts a fresh session, picks the next task from the active plan, implements it, tests it, commits, and exits. Fresh context per iteration is the key design principle — avoids context window degradation.

Every N worker iterations (default 5) and once at the very end, the loop runs an audit iteration instead of a worker iteration. The auditor spawns parallel subagents to review recently completed work against the plan and codebase, triages the findings, and injects follow-ups into the plan as new [ ] tasks. It never fixes code itself — the next worker iteration picks the audit tasks up normally.

The user creates the plan (via the planning skill or manually). The loop only implements — but the agent can update the plan when it discovers new work, bugs, or needed refactoring.

Bundled Templates

These templates ship with the skill. Copy them into the target repo under loop/ and gitignore that directory. The templates are designed to be project-agnostic — most projects need zero changes, some need a tweak to the audit prompt's checklist.

| Template | Purpose |