paidv0.1.0unverified
Loop Patterns — the four loops of loop engineering, installable
"Loop engineering" is the defining AI-engineering skill of 2026 (swyx's
framing; HN's most-requested tooling). This pack ships the four loop types as
four skills you compose — each a one-page contract, not a framework.
Quick start
Preview (free): skills/verify-loop/ is included here — copy into
.claude/skills/, then: `/verify-loop draft the pricing page copy, rubric:
clear, honest, no hype`.
Full bundle ($49): all four loops + the composition guide.
What you get
| Loop | Contract | In preview? |
|---|---|---|
/verify-loop | generate → judge against explicit rubric → revise → repeat until pass or plateau | ✅ free |
/agent-loop | do → check via command exit code → fix → repeat (generalized test-loop) | paid |
/event-loop | watch a trigger (file, cron, webhook log) → act → acknowledge → re-arm | paid |
/climb-loop | measure baseline → try one variant → keep if metric improves → log attempt memory | paid |
COMPOSING.md | when to stack which loops (verification inside agent inside event is the workhorse) | paid |
The one primitive
Every loop is the same four decisions: **what repeats, what decides (never the
model's opinion — a command, a rubric, a metric), what counts as stuck, and
what survives between iterations.** The four loops are the four useful answers.
Pairs with agent-loops-starter (free) and
window-surfer's /goal for time-budgeted runs.
Proof
Unbenchmarked — v0, not yet for sale. First benchmark: /climb-loop on a
prompt-optimization task, measuring metric improvement per iteration vs manual.
Changelog
- 0.1.0 (2026-07-04) — initial release, unverified.
One email when a pack ships or gets re-benchmarked on a new model. No noise.