← ClaudeAtlas

loop-templatelisted

Use when creating any new automated loop, scheduled task, or launchd job, or when the user says 'new loop', 'build a loop', 'add an automation', or 'scaffold a scheduled task', so the loop inherits the maker/checker pattern.
courtneyr-dev/agent-skills · ★ 0 · Data & Documents · score 75
Install: claude install-skill courtneyr-dev/agent-skills
# Loop template How to build a loop that's honest. A loop is a recurring task that finds its own work, does it, and verifies it — without you being the cron or the only checker. ## Step 0: the "when NOT to loop" gate (run this first) Do not build the loop if any of these is true: - **High slop-risk AND low verification confidence.** If the output is hard to auto-verify and a bad run pollutes a vault/repo/channel, keep it manual. - **The follow-up is a <5-minute UI click, or it needs the user's eyes anyway.** Automating a thing you still has to review by hand just adds a step. - **The output is taste, not facts.** Prose, atom-writing, social copy, design — loops can *draft* these into a "NEEDS REVIEW" file, but never auto-publish and never treat them as done. If it passes the gate, continue. ## Step 1: name the five blocks Write these down before any code: 1. **Automation** — what fires it? (cron schedule, launchd, hook, or manual) 2. **Maker** — the produce step. What does it create or change? 3. **Checker** — the verify step. See the rule below. This is the part that makes "done" mean something. 4. **Memory spine** — the file on disk that holds state between runs. Without it the loop can't tell "broken once" from "broken three weeks running." 5. **Connectors** — which MCP tools / CLIs it touches (gh, wp-cli, Jira, Beeper, Obsidian). ## Step 2: the blind-checker rule (non-negotiable) **The checker re-derives findings from source. It does NOT re-read the maker's outp