Andersonlimahw Prompts — Feedback and quality loops
Recipes to turn "I think it works" into evidence. Golden rule of this track: one loop driver at a time — chain loops in sequence, never two owning the same cycle.
R1 — API validation loop
Skills: api-test-loop + verification-before-completion
When: new or changed endpoints; you want proof via CURL, not opinion.
Run api-test-loop on endpoints <list or route prefix>. Validate
REST practices, edge-case inputs (nulls, negatives, overflow,
injection), security, and output format via CURL. Log every
finding in findings.md with severity and a reproduction curl.
Apply surgical fixes and re-validate until all CRITICALs are
closed. Close out with verification-before-completion.
Success: findings.md with every CRITICAL closed, each with the
before/after proof curl.
R2 — Exploratory browser QA
Skills: chrome-qa-loop
When: a live web app needs exploratory eyes, not an E2E suite.
Run chrome-qa-loop on app <url>, using screen manifest
<path/config>. Explore each screen with the doc's context, produce
one markdown report per finding (severity, repro steps, evidence).
Don't fix anything yet — triage only.
Success: one report per finding, triageable without opening the
app; zero duplicate findings.
R3 — Error-fix loop
Skills: error-fixer-loop
When: a build/test/typecheck broke and you want a fix + anti-regression rule, not a band-aid.
Captured this failure: <paste exact stack/error>.
Run error-fixer-loop: investigate root cause → minimal fix →
verify with the same command that failed → persist the
anti-regression rule for this error pattern.
Success: verification command green + rule persisted; the diff
touches only what's necessary.
R4 — Metric optimization loop
Skills: karpathy-loop
When: an objective metric exists (latency, bundle size, tokens, coverage) and you want autonomous improvement cycles.
Run karpathy-loop with target: improve <metric> from <baseline> to
<goal> in <file/area>. Each cycle: hypothesis → minimal change →
measure with <measurement command> → keep or revert. Stop at <N>
cycles or at the goal.
Success: metric at goal (or best result after N cycles) with a log
of every cycle and a reproducible measurement.
Recommended chaining
chrome-qa-loop (find) → bug-diagnostics (diagnose the tricky finding) → error-fixer-loop (fix with a rule). Sequence, never simultaneous — each stage hands off an artifact to the next.