start_skill

Overview

This skill runs an autonomous iteration loop that executes multi-step tasks until the spec is satisfied or configured limits are reached. It can also explain how the Loop mechanism works and what checks are enforced before, during, and after iterations. Use it to persistently implement features, fix problems, or verify success criteria with strict state tracking and safety gates.

How this skill works

The Loop inspects the incoming spec, scores clarity across five dimensions, and determines shape (tool vs colleague vs intent) to decide how to run. It decomposes the task into concrete steps, creates a structured state file, and performs repeated iterations that announce progress, execute one logical step, verify criteria, write .loop/state.json, and update task state until all criteria are met or a circuit breaker pauses the run. Mandatory announcements and a final <loop-complete> marker ensure the stop hook can detect completion.

When to use it

• Start an autonomous run for multi-step engineering work (refactors, feature implementations).
• Persistently attempt fixes until tests and explicit success criteria are met.
• Run repetitive verification and remediation cycles where state must be recorded after every iteration.
• Resume a paused workflow with explicit budget or iteration overrides.
• Ask how Loop works or request help deciding if the task is tool-shaped or colleague-shaped.

Best practices

• Score the spec with the five-dimension rubric before starting; if score <5, clarify intent first.
• Extract explicit success criteria and list them in .loop/state.json so verification is automatable.
• Decompose work into concrete steps and enforce one logical unit per iteration for predictable progress.
• Use conservative default budget and iteration caps; rely on pause hooks rather than killing the run.
• Respect rigid announcement formats and write .loop/state.json after every iteration so stop hooks and recovery can function.

Example use cases

• /loop refactor all API routes to validation middleware, continue until tests pass and routes updated.
• /loop fix flaky CI failing specific test suite until all tests are green or budget exhausted.
• /loop implement full feature end-to-end with state persisted each iteration and progress announcements.
• Ask “what is loop” to receive the checklist, rubric, and iteration protocol before starting.
• Resume a previously paused loop with /loop continue --budget=50 to extend budget and continue iterations.