jrpease/agent-bootstrap/verify

Verify

A task is not complete because code was written. Verification scales with what changed — run the narrowest set that gives real confidence, and never report a check you did not run.

1 · Pick the ladder

What changed decides what runs. Work down; stop when the remaining rungs cannot fail for this change.

Discover commands from the repo (package.json scripts, Makefile, CI config, the project's agent instructions) rather than guessing. No test infrastructure is a fact to report, not a rung to fake.

2 · The browser loop (frontend changes)

Typecheck proves the code compiles. Only a browser proves the change works.

Launch the app (dev server or built output; where a `run` skill exists it
knows the per-project command, otherwise read it from package.json/Makefile)
   ↓
Open the affected page at a real http:// URL
   ↓
Look at it — screenshot desktop (1440) and mobile (375)
   ↓
Exercise the changed interactions (click, type, drag — what the change touches)
   ↓
Read the browser console — zero new errors or warnings
   ↓
Fix → repeat until clean

• A design reference exists → compare against it. Figma frame, DESIGN.md,

or an existing page in the same system. Judge the render against the reference, not the DOM against the code.

• Responsive states: every viewport the change plausibly affects, minimum the

two above.

• Pages on virtual scroll (Lenis/Locomotive): try

window.scrollTo({ top, behavior: 'instant' }) first and assert the scroll progress you expected. Stepped wheel events overshoot, because Lenis keeps gliding after each one. Fall back to wheels (or click, then PageDown) only when the page ignores scrollTo and captures show one frame five times.

• Headless Chrome ignores the autoplay policy, so "audio waits for a gesture"

can't be verified headless.

3 · The heavy harness (marketing/campaign pages)

Pages built under studio-design are verified by its loop (verify.md there) — do not duplicate it here. But its harness is reusable for any page that needs recordings or performance evidence:

Run these from the studio-design skill's reference/ directory — resolve its path from wherever your skills are installed, then:

node <studio-design>/reference/record.mjs http://127.0.0.1:<port>/ --out captures
node <studio-design>/reference/perf.mjs   http://127.0.0.1:<port>/          # LCP/CLS/INP + frame budget
npx impeccable@3.5.0 detect http://127.0.0.1:<port>/ --json                 # tell detector

Never run `impeccable install` — in any form; detect needs none of it. Perf bars when measured: LCP ≤ 2500ms · CLS ≤ 0.1 · INP ≤ 200ms. A zero is a missing measurement, not a pass.

Glance freely (screenshot, look, fix); cap expensive full rounds at 3 and report what remains rather than looping past the cap.

4 · Report

Implemented:
- <change>
 
Verified:
- <check> — <real result, numbers not adjectives>
- browser: <page> · <interactions exercised> · console clean · 1440 + 375
 
Not verified:
- <what, and why>          # absent infrastructure, unreachable state — named, never silent
 
Remaining risk:
- <concern, or "none">

An honest "not verified: X" is a deliverable. A claimed check that never ran is the one unforgivable output of this skill.