← ClaudeAtlas

loop-designlisted

Loop engineering skill: turn a recurring intent into a deployed, safe, self-expiring automation loop (headless Claude/Codex sessions on a schedule). Use for ANY of: designing recurring automation ("design a loop", "run X automatically every night", "scheduled job", "advance a project overnight", "burn quota", "automate this workflow"); choosing the mechanism ("cron or /loop", "session-only or persistent", "cloud routine or local"); writing unattended prompts ("how do I write a headless prompt", "unattended", "is auto-commit safe"); scheduling ("how do I schedule it", "avoid rate limits", "run while I sleep"); reviewing/debugging an existing loop ("the loop didn't run", "the morning report is low quality", "it deadlocked", "review the loop's output"). Also triggers on: cron, crontab, launchd, nightly, scheduled, polling, unattended, headless.
colossus8756/loop-design · ★ 0 · AI & Automation · score 73
Install: claude install-skill colossus8756/loop-design
# Loop Design — Unattended Automation Design Skill > Reference implementation: `~/automation/` (a nightly loop system: > run-loop.sh + prompts/ + logs/ + reports/). Keep runners and prompts outside > iCloud-synced folders (see §6 for why). When designing a new loop, prefer reusing > this pattern over building from scratch. ## 1. What This Skill Does Turn "I want an AI to do X regularly/repeatedly" into a loop system that is **bounded, gated, self-expiring, and has a clear exit**. It produces four artifacts: a mechanism decision → a wrapper script → one prompt file per loop → a schedule (crontab or equivalent). ## 2. Run Protocol ``` 1. Classify — Which kind of use is this? (§3 five types). Split mixed cases into separate loops. 2. Pick a mechanism — Choose one of the four in §4; the wrong mechanism is the most expensive mistake. 3. Map the clock — First mark the user's unavailable windows (sleep/work); loops only fill dead time. Leave live time for the user's own interactive work (or you fight them for quota). 4. Write the prompt — One standalone prompt file per loop; self-check against each of the 8 rules in §5. 5. Add safety layers — date guard + one lock per loop (with stale-lock cleanup) + git isolation (§6). 6. Schedule — Interval ≥ one round's duration; avoid the top/half of the hour; put wrap-up loops last, so their human-readable report lands before the user's next live window starts. 7. Smoke — Validate the whole