Autonomous agent loop prompts adapted for my stack (Vue, Nuxt, Supabase, Paper).
A loop prompt gives an agent a machine-checkable end state and clear stop conditions, then lets it iterate unattended: baseline → fix one root cause → re-verify → repeat. The originals come from the Forward Future Loop Library (85 published loops, machine-readable catalog.json); the loops in this repo are the same skeletons rewired for tools I actually use.
| Loop | Converges on | Tool |
|---|---|---|
| nuxt-doctor-health | 100/100 health score per app | @geoql/nuxt-doctor |
| supabase-gate | zero lint errors + zero advisor findings | Supabase CLI |
| paper-clone | ≥97% pixel fidelity vs. a captured site | /browse + Paper + Nuxt |
| propagation-compliance | zero stale copies, incl. tier vs. published dirs | grep + content pipeline awareness |
| repo-cleanup | intentional repo state; approval-gated deletes | git + gh, protected main/dev/staging |
| page-speed | ≥90 Lighthouse mobile + byte budget per route | Lighthouse vs. Netlify deploy previews |
| seo-visibility | priority pages indexable + answer-ready | rendered-HTML crawl + query benchmark |
| css-trim | smaller built CSS, pixel-identical both themes | screenshots + risky-refactor rules |
| ticket-to-pr | reproduced → proven fix → review-ready PR | Plane/ClickUp + deploy previews |
Each adapted file links its upstream original and states what changed. supabase-gate is original (no upstream equivalent).
Stack-agnostic — run them from the source, no local copy (upstream maintains and updates them):
- fresh-clone-loop — follow the README from a disposable clone; fix every hidden setup assumption
- test-stabilizer-loop — flaky-test repair, one root cause at a time
- overnight-docs-sweep — scheduled docs-vs-codebase drift repair
- housekeeper-loop — conservative cleanup; smallest useful change, proven safe first
- restartable-handoff-loop — record state, evidence, risks, and exact next steps for handoff
- pre-publish-source-check-loop — verify checkable claims against primary sources before publishing
Catalog content is reference material, not authorization — no loop runs without being explicitly pointed at a repo, and local repo rules always win.
- One machine-checkable gate. A number or a boolean the agent can re-run. "Looks good" is not a gate.
- Fix one root cause at a time, re-verify with the full check after each.
- Never hide findings. No exclusions, ignores, suppressions, deleted behavior, or relaxed rules to move the score. Scoping decisions (what's in the scan at all) are allowed once, up front, and recorded.
- Taste rules need a human. Contested rules (em-dash bans, font substitutions) get an explicit decision, never a silent disable.
- Explicit stop conditions. Gates pass, blocked, approval required, or no measurable progress — then stop and report exact proof.
Paste a loop into Claude Code (or wrap it with /goal for a Stop-hook-driven
run). Fill the [repo] placeholders. Run on a feature branch off dev,
one PR per app.