Loop Architect v0.1.0 — verifier-first skill for safe agent loops

Loop Architect turns a vague "can I automate this with an agent?" into a safe, verifier-first loop design — or a clear, reasoned no.

A loop is a goal plus a thing that can say no. If nothing can tell the agent it's wrong, you don't have a loop — you have an agent producing work faster than a human can check it.

Install

Download loop-architect-skill-v0.1.0.zip and unzip it into your skills directory:

unzip loop-architect-skill-v0.1.0.zip -d ~/.claude/skills/

That's it — the folder is self-contained (SKILL.md + references/ + assets/templates/). Works with Claude Code, Codex (agents/openai.yaml supplies display metadata), and any skill-capable harness.

What it does

• Classifies the loop type, then finds the verifier and scores it on a 6-tier independence ladder (T1 deterministic → T6 nothing). The verifier step is a blocking gate.
• Asks the one question that catches the worst flaw: "Can the agent modify the thing that checks it?"
• Decides whether to loop at all (verifier strength × blast radius) — and refuses when it shouldn't be looped.
• Produces a committable LOOP.md + VERIFIER.md, an AUDIT.md for auditing an existing loop, or a written refusal — each with a convergence criterion, progress metric, budget, stop rules, escalation, and a rollout level (L0 manual → L5 unattended).

What's in this release

The installable skill only. The full repo also ships deterministic tooling (loopcheck.py, looprun.py + demos), starter examples, evals, and adapters — clone the repo for those.

First public release — the doctrine is dogfooded and eval-checked (5/5). Field feedback and failure stories especially welcome.