Loop

Loop

Use this skill to run supervised autonomous loops without drifting from the user's goal.

Loop Contract

Before starting, define:

• objective: the concrete target state
• cadence: how often to check or iterate
• max rounds or time limit
• stop conditions
• rollback or pause conditions
• verification signal

If the user did not specify limits, choose conservative defaults:

• max rounds: 5
• pause on destructive operations, unclear requirements, credential exposure, or repeated identical failure
• verify after every round with the closest reliable check

Iteration Pattern

For each round:

1. Inspect current state and last failure.
2. Select the smallest next action.
3. Execute the action.
4. Run the agreed verification.
5. Record status:
   • LOOP_CONTINUE: progress made, more work remains
   • LOOP_PASS: objective met
   • LOOP_BLOCKED: needs user input or external resource
   • LOOP_STOP: unsafe or no useful progress

When To Stop

Stop immediately when:

• the objective is met
• the same failure repeats after two different fixes
• a security or privacy issue appears
• changes would exceed the requested scope
• tests require unavailable external services and no safe mock exists
• continuing would require destructive file, database, or git operations not requested by the user

Status Report

For long loops, provide compact updates:

• round number
• action taken
• verification result
• next action or stop reason

Final report should include:

• final status
• changed files or artifacts
• verification commands and results
• remaining risks or manual follow-up

Related Skills

Combine with:

• verification-before-completion before claiming completion
• systematic-debugging when failures are unclear
• requesting-code-review for substantial code changes
• simplify when the loop is fighting unnecessary complexity