walkthrough-review

Walkthrough Review Skill

Use this skill when the user asks for a recorded product walkthrough, demo GIF/video, Gemini video understanding review, README media proof, or a "prove the UX is good" loop. The goal is not a hero screenshot. The goal is a reproducible workflow:

feature spec -> browser capture -> annotated render -> Gemini judge -> defect fixes -> README proof

Non-Negotiable Bar

Do not claim a walkthrough is good, optimal, README-ready, or production-quality unless all of these exist:

• A browser-recorded walkthrough artifact: docs/walkthroughs/<feature>.mp4 or .gif.
• A model-assisted review artifact from Gemini: docs/eval/gemini-media-judges/<run>/summary.md.
• A clear defect disposition: P0/P1 fixed before publishing; P2 either fixed or documented.
• A README or docs link that says what the clip proves and what it does not prove.

The judge should compare the interaction quality against polished professional tools such as Notion and Linear: calm hierarchy, obvious active state, dense-but-readable information, low step count, and no ambiguous mode changes.

Primary Command

Run the NodeRoom-compatible wrapper:

npm run walkthroughs:review -- startup-diligence-war-room --ui-review

The wrapper delegates to the reusable CLI package:

npm run walkthrough-review -- startup-diligence-war-room --ui-review

Useful variants:

# Recapture and rejudge the two flagship startup clips.
npm run walkthroughs:review -- startup-diligence-war-room startup-diligence-live-join --ui-review

# Rejudge existing media only, without recapture/render.
npm run walkthroughs:review -- startup-diligence-war-room --skip-capture --skip-render --ui-review

# Use a specific base URL, e.g. local dev or production.
npm run walkthroughs:review -- startup-diligence-war-room --base=http://127.0.0.1:5178 --ui-review

# Intentionally judge already-rendered media even if a fresh capture skipped.
# This is advisory only; never use it for fresh evidence claims.
npm run walkthroughs:review -- startup-diligence-war-room --allow-skipped --skip-render --ui-review

The wrapper writes a run manifest to:

docs/eval/walkthrough-review/<run-id>/README.md

and the media judge writes:

docs/eval/gemini-media-judges/<run-id>/summary.md

Reusable CLI And MCP

The reusable package lives at:

packages/walkthrough-review-cli/

Project-specific wiring is in:

walkthrough-review.config.json

The config owns the actual capture/render/review commands; the package owns argument parsing, command templating, optional local env loading, capture-manifest validation, media picking, and JSON/Markdown report generation.

CLI form:

npm run walkthrough-review -- startup-diligence-war-room --ui-review

MCP form for coding agents:

npm run walkthrough-review:mcp

The MCP server exposes walkthrough_review_run and calls the same runner as the CLI. Do not maintain separate browser-capture or review logic in MCP.

Capture Rules

• Prefer a fresh room for live captures unless the spec explicitly uses memory mode.
• The wrapper fails by default if a requested feature is skipped or captured with zero segments. Use --allow-skipped only when intentionally judging old media.
• Show every state: empty/before state, cursor target, loading/streaming, result.
• For collaboration, use one browser context per persona. Shared localStorage silently reuses the first persona and invalidates the proof.
• Keep captions short and outcome-oriented. Captions explain why the step matters, not just which button was clicked.
• Make active mode switches visible. Public/private, spreadsheet/note/wall, and review states need a visible active tab or transition beat.
• Do not hide workflow latency. If work streams or waits, capture the wait honestly.

Gemini Review Rules

Use gemini-3.5-flash by default:

npm run media:gemini-judge -- --only startup-diligence --include-ignored
npm run ui:gemini-review -- --media=docs/walkthroughs/startup-diligence-war-room.mp4

Interpretation:

• publish: acceptable as README/demo evidence.
• fix-then-publish: fix P0/P1 before making it primary evidence.
• rework: rebuild the walkthrough or the product surface before using it.

The media judge is evidence for demo quality only. It does not replace Convex typecheck, browser E2E, provider ladder, privacy, load, or cost gates.

Iteration Loop

1. Run npm run walkthroughs:review -- <feature> --ui-review.
2. Read docs/eval/gemini-media-judges/<run>/summary.md.
3. Fix every P0/P1. Fix cheap P2s that improve legibility or mode clarity.
4. Re-run the same command with a new run id.
5. Update README only after the review artifact exists.

Files To Edit

• Walkthrough tape: scripts/walkthroughs/specs.ts
• Browser capturer: scripts/walkthroughs/capture.ts
• Renderer: scripts/walkthroughs/render.ts
• Full wrapper: scripts/walkthroughs/review.ts
• Gemini media judge: scripts/gemini-demo-media-judge.ts
• Gemini production UI rubric: scripts/gemini-ui-review.ts
• README proof section: README.md
• Reusable CLI/MCP runner: packages/walkthrough-review-cli/
• Project CLI config: walkthrough-review.config.json

Commit Checklist

Before committing:

npm run typecheck -- --pretty false
npx tsc --noEmit --project convex\tsconfig.json --pretty false
npm test -- --run tests/contentFluency.test.ts tests/improvementArtifacts.test.ts
npm run build

Track GIFs and docs. MP4s are intentionally ignored as regenerable local evidence.