designing-loops

A Claude Code skill that picks the right loop type for agentic work and writes stop conditions that actually hold.

The four-type taxonomy (turn-based / goal-based / time-based / proactive) comes from the Claude Code team's article Getting started with loops. The rest comes from running real multi-hour autonomous loops and auditing what they claimed afterwards:

• Write the stop condition first — a number beats an adjective, and every loop needs a dual stop (success OR bounded failure). Vague done criteria is how a loop turns into a space heater.
• Guard against goal redefinition — long loops under pressure quietly re-scope "done". Pin the original metrics where the loop can't edit them.
• Heavy tier: give the loop a runbook file — for overnight/autonomous runs, durable state belongs in a committed file (goals table, tiered guardrails, work queue, ledger), not in the prompt. Template included in references/runbook-template.md.
• Audit after the loop — the final report is a claim, not a fact. Re-measure the headline numbers yourself.

Install

git clone https://github.com/edwluo/designing-loops.git ~/.claude/skills/designing-loops

That's it. Next time you tell Claude Code "babysit this PR", "run this overnight", or "fix all 47 errors", the skill kicks in — it classifies the loop, writes the ready-to-run prompt with stop conditions and budget included, and for heavy autonomous runs generates a full runbook file.

It also knows when to say no: ask it to "loop until the README reads perfectly" and it will tell you that's one turn of work plus a review pass, not a loop.

What's inside

License

MIT