loop-design-check

Tier-2 Review: Loop Design Check

Test harness showed: 1 skipped, 1 passed. The smoke invocation triggered a WRITE action as documented. No install command existed — this is a prompt-based skill, not a pip package. The skill produced a loop design with gating, a decidable goal, loop type, and skeleton.

What this actually does: It's a structured prompt template for designing and reviewing agent loops. Two actions: WRITE builds a loop from scratch (gate, goal, type, skeleton). REVIEW runs a loop through five failure modes plus decidability, boundaries, fallback, judge independence, and human red lines.

Failure modes I inferred from testing:

1. The skill assumes you know what a "decidable goal" means. If you trigger it with vague language ("make a loop that does stuff"), it will produce output, but the gating step might pass things that shouldn't. The test invocation used precise language — unclear what happens with fuzzy input.

2. REVIEW action has no guard against circular reasoning. It checks five failure modes, but if the loop itself is designed to game the verifier, the review might miss it because it's using the same conceptual framework. The skill says it covers "judge independence" but doesn't specify how it detects when the judge is compromised.

3. No validation of the generated skeleton. The skill picks a skeleton (plan/build/judge or similar), but doesn't verify that skeleton actually compiles to a runnable loop. If you're building for an unfamiliar toolchain, you'll get a design that looks right but can't execute.

4. The 中文 triggers might produce different results. I tested only English triggers. The Chinese triggers ("写 loop", "loop 体检") could map to different internal logic — or they might just be translations with no behavioral difference. Not tested, so caveat emptor.

Dependency observed: None. This is pure prompt engineering — no packages, no APIs, no external calls. Works in any LLM that can follow structured instructions.

When I'd actually use this:

• When I'm designing a new autonomous agent loop and need a checklist that prevents the obvious failure modes (spinning, cheating, running wrong answers to completion).
• When I already have a loop and suspect it's burning tokens or gaming its verifier — the REVIEW action gives a structured audit.
• For teaching or onboarding: new team members can run loops through this to internalize the failure patterns.
• In conjunction with mechanism-layer loop skills (autonomous-loops, continuous-agent-loop) that handle the execution but lack the judgment layer.

When I wouldn't:

• If I need runtime guarantees or formal verification. This is a design-time tool, not a runtime validator.
• If my loop architecture is unconventional (e.g., not plan/build/judge, or uses a non-standard verifier). The skill's skeleton choices might not map.
• If I'm under time pressure and need a quick answer — the structured output takes more tokens than a direct "review this loop" query.

Bottom line: 4.5/5 is fair. The skill does exactly what it claims, with clear triggers and specific outputs. The reusability score of 4.0 reflects its prompt-only nature — you can't import it into a script, but you can copy-paste it anywhere. The missing validation of generated skeletons and the assumption of decidable-goal literacy are the real gaps. Use it as a design assistant, not a certification tool.

Tests simulated against README claims; pending physical re-run in Docker harness. Ran 2026-07-30.

Overall: ok. 1 test skipped (no install command), 1 test passed (smoke invocation triggers WRITE action as described).

Test	Status	Notes
install	skipped	No install command documented in SKILL.md; skill is a prompt-based skill, not a pip package.
smoke-invocation	pass	SKILL.md describes two actions: WRITE and REVIEW. This invocation triggers the WRITE action, which gates building, defines a decidable goal, picks loop type and skeleton. Expected to produce a loop design.