ui-recorder

ui-recorder Skill

Orchestrates the 5-stage pipeline: plan → record → narrate → edit → render.

Workflow

Stage 0 — Plan-Build (verifies selectors against real DOM)

1. Slugify the user's description → workspace slug.
2. Run: node "${CLAUDE_PLUGIN_ROOT}/scripts/plan-build.mjs" <slug> -- <user-prompt>
3. Read stdout:
   • BYPASS → user already placed browse-plan.json (e.g. from test-fixtures/). A synthetic plan-report.json was written with source: "user-provided" for all actions. Skip to Stage 1.
   • NEEDS_AGENT → workspace is empty. Dispatch the plan-builder agent (Task tool, subagent_type plan-builder). Pass it the workspace path, user prompt, viewport, and playbackRate. The agent writes browse-plan.json, plan-report.json, and narration.txt into the workspace.
4. On agent failure (errors in stderr): surface the code + hint. If it reports PLAN_CHROMIUM_MISSING and no browser MCP was available, offer the user: (a) run setup.sh to install chromium, (b) proceed with the agent's partial plan, marking unverified actions as blind manually, (c) edit plan manually.

Stage 1 — Review (Gate 1)

5. Read workspace/{slug}/browse-plan.json and plan-report.json. Present to the user:
   • Summary: N actions across M URLs, estimated output ~Xs at the configured playbackRate.
   • Per-action line: #N <action> <label> — selector: <sel> [confidence: <level>, source: <tier>] + short rationale.
   • If any action has source: "blind", flag with a warning icon and show the warning field.
6. Gate 1: "Proceed? (yes / edit selector N / drop action N / regenerate)". Wait for explicit approval.
7. On "edit" — update browse-plan.json and the matching action in plan-report.json (change source to "user-provided"), then re-prompt.
8. On "regenerate" — ask the user for a refined prompt, then rerun Stage 0 from step 2.

Stage 2 — Record

5. Invoke node "${CLAUDE_PLUGIN_ROOT}/scripts/record.mjs" <plan.json> <slug>.
6. On failure, surface the code + hint from UIRecorderError. Suggest the fix verbatim.
7. On success, report: recording.mp4 path, moments.json summary (N actions captured, total duration).

Stage 3 — Narrate

8. Check workspace/{slug}/narration.txt — Stage 0's plan-builder agent writes a draft when it runs. Present the draft (if present) alongside the prompt: "Narration? (a) use agent draft, (b) write me a script, (c) auto from action labels, (d) silent". If no draft exists, offer (b)/(c)/(d) only.
9. For (a): keep the existing narration.txt. For (b): prompt for the script. For (c): join all label fields into a neutral sentence per action. For (d): delete narration.txt and skip to Stage 4.
10. Write workspace/{slug}/narration.txt (unless (a) kept it as-is or (d) deleted it).
11. Invoke node "${CLAUDE_PLUGIN_ROOT}/scripts/narrate.mjs" <slug>.
12. If TTS_OVERFLOW: auto-rewrite was already tried. Ask user to shorten manually.

Stage 4 — Edit (Gate 2)

13. Dispatch the edit-planner agent with the workspace path. It reads moments/alignment and writes camera-config.json.
14. Gate 2: Show segments list, total output duration (with 4× speedup), camera movement summary. Ask "render now? (yes / adjust)".
15. If user wants adjustments, let them specify zoom levels, segment bounds, or crop strategy → update camera-config.json manually → proceed.

Stage 5 — Render

16. Invoke node "${CLAUDE_PLUGIN_ROOT}/scripts/render.mjs" <slug>.
17. On success, report: demo.mp4 path, file size, duration.
18. Offer adjustments: re-render with different zoom, aspect ratio (v0.1: only 16:9), or brand.

Natural-language config updates

Recognize phrases like:

• "use voice X" → update config.json tts_voice
• "faster pace" / "slower pace" → update browse-plan.json playbackRate (default 2)
• "silent" → skip Stage 3

Write config via loadConfig/updateConfig from scripts/lib/config.mjs.

Caching

Never invoke an earlier stage if its cache sidecar is valid. The scripts handle this internally via .sha256 sidecars. If the user says "re-record", delete ${workspace}/recording.mp4.sha256.

Error recovery

When a script exits non-zero, read ${CLAUDE_PLUGIN_DATA}/logs/last-error.log via Read, extract the UIRecorderError code + hint, and present to the user. Do NOT auto-retry without asking (policy set in design §6.3).

What v0.1 does NOT support (redirect politely)

• Recorder: local-headed, steel — redirect to local-headless
• TTS: elevenlabs — redirect to piper
• Aspect ratios other than 16:9
• Brand extraction from screenshots
• Authenticated flows, iframes