kookr-demo-recording

name: kookr-demo-recording description: How to record, verify, export, and publish the Kookr demo video with optional TTS narration audio keywords: demo, video, recording, tts, text-to-speech, playwright, narration, audio, webm, mp4, vertical, shorts, gif, srt, subtitles, thumbnail, youtube, publish, regenerate demo, check mode related: playwright-e2e-patterns, e2e-agent-testing

Demo Recording

When to Use

When you need to generate, verify, or regenerate the Kookr demo video and its distribution formats, with or without voice narration.

Quick Reference

ALWAYS check scenario alignment first

The recording drives the real current frontend (built by pnpm build, served by the E2E test server), so the video always shows the current UI by construction. What drifts are the selectors and flows in the demo script. Before any recording — and after any significant frontend change — run:

pnpm demo:check

This replays the entire scenario headless with all holds capped to 150ms, no TTS, no video, no exports (1 min). It fails loudly on any selector or flow drift and ends with CHECK PASSED when the script is fully aligned with the current UI. Never record without a passing check.

Always prefer narrated video

The .env file in the repo root has HF_TOKEN and KOOKR_TTS=true. Copy it to the worktree if needed, then record with TTS:

cp $HOME/git/kookr/.env .env  # if in a worktree
pnpm demo:record                    # reads .env automatically

Only record silent if TTS is genuinely unavailable (Docker down, no HF token).

Silent video (fallback only)

KOOKR_TTS= pnpm demo:record

With external TTS service (already running)

KOOKR_TTS_URL=http://localhost:8004 pnpm demo:record

Opening hook variants (A/B testing)

The first narration line is selectable — everything else stays identical, so variants are cheap to regenerate and A/B test on social:

DEMO_HOOK=A pnpm demo:record   # default: "Running five coding agents sounds like leverage…"
DEMO_HOOK=B pnpm demo:record   # "Your AI agents are fast. Supervising them is the part that does not scale."
DEMO_HOOK=C pnpm demo:record   # "This is what running five AI coding agents actually looks like…"

How It Works

Pipeline

1. TTS setup (if enabled): starts Docker container (tts/docker-compose.yml), generates WAV clips for each narration line via POST /synthesize
2. Server start: launches E2E test server with FakeTerminalBackend + FakeTerminalBridge (no tmux needed)
3. Playwright recording: drives the UI scenario with recordVideo, injects captions via DOM. The video opens directly on the cold-open terminal grid (pain first — social retention is decided in the first 3 seconds); the logo appears only on the closing CTA card.
4. Audio sync: holdTime() ensures each caption holds long enough for the speech clip to finish (clip duration + 500ms padding)
5. ffmpeg merge: combines silent video + timed audio clips using adelay filters
6. Music bed (optional): first audio file in demo/assets/music/ is looped, faded, and mixed at low volume under the narration
7. Subtitles: kookr-demo.srt is generated from the same narration timeline (upload with the video so YouTube can auto-translate captions)
8. Export matrix: every distribution format is derived from the master WebM (see Output below)
9. Cleanup: stops TTS container, removes temp files

Key files

Terminal content modes

• instant — dumps all content at once. Used for blocked/stopped agents (permission block, needs input). Terminal appears frozen.
• streaming — sends line-by-line at configurable speed. Used for healthy running agents. Conveys "this agent is busy."

Narration scripts

Defined in NARRATIONS object in demo/record.ts. Each key maps to a tracker.mark() call in the scenario. To change narration text, edit the NARRATIONS object. Opening hook lines live in HOOK_VARIANTS (selected via DEMO_HOOK).

Voice selection

Default: Matilda (/app/voices/matilda.mp3). Override with:

TTS_VOICE=alba pnpm demo:record   # Built-in voice (no HF_TOKEN needed)

Built-in voices (no token): alba, marius, javert, jean, fantine, cosette, eponine, azelma

Output — export matrix

One recording produces every distribution format:

Music bed

Drop one license-free track into demo/assets/music/ (mp3/wav/m4a/ogg/flac — first file alphabetically is used). It is looped to the video length, mixed at ~9% volume under the narration, faded in 2s / out 3s. No file → narration-only (a one-line log says so). Only applied to narrated recordings.

TTS service lifecycle

Mirrors the STT pattern exactly:

• KOOKR_TTS=true → auto-start Docker container on server start, auto-stop on shutdown
• KOOKR_TTS_URL → use external service (skip Docker)
• Non-fatal: TTS failure doesn't break the server or recording (falls back to silent)
• Manager: src/server/tts-manager.ts (mirrors src/server/stt-manager.ts)

Emoji Rendering

Recording requires fonts-noto-color-emoji on Linux / Apple Color Emoji on macOS — verified by demo/lib/preflight.ts at startup.

Fake Data Requirements

The demo uses FakeTerminalBackend — no real agents run. This means features that depend on real infrastructure need fake data injected:

Rule: every caption claim must have corresponding injected data. If you say "cost tracked in real time", the TopBar must show a nonzero dollar amount.

UI behaviors that have bitten the scenario before

• Onboarding tour: suppressed via ?onboarding=0 URL param (version-proof) + the kookr:onboarding:seen-v2 localStorage key. If the tour overlay intercepts clicks, the storage key version bumped — update both.
• Project-filter sync: selecting a task switches the project filter to that task's project. After triaging a webapp finding, api-service agents are filtered out — click the all-projects chip before showcasing them.
• Findings auto-scroll: a new finding scrolls the list to top, which can swallow a click landed mid-scroll. selectFindingByText retries once.
• Achievement toasts: pop at uncontrolled times bottom-right and read as noise to first-time viewers — suppressed via injected CSS for the entire recording.
• Launch cwd validation (RFC F12): the production task-creation API rejects nonexistent working directories with invalid_cwd. The e2e test server no-ops the check (#871), but seeded agents use real temp dirs ($TMPDIR/kookr-demo-cwd/..., created in record()) so the demo never depends on that no-op staying in place.
• Provider badge is in the collapsed "Details" popover: .detail-agent-provider resolves but is hidden until the .detail-meta-menu > summary is clicked. Open the popover on screen before zooming on the Codex badge.
• Terminal content binds at WebSocket connect: calling set-terminal-content on an already-open terminal does nothing visible. To show an agent "resuming", set the new streaming content first, then (re)select the task so a fresh terminal connection picks it up.
• The worktree .env carries real Telegram credentials: the demo server is forked with KOOKR_REMOTE_CHAT_DISABLED=1, otherwise injected demo alerts send actual Telegram notifications mid-recording.
• The startup curtain must be in the served HTML, not script-mounted: Vite's module scripts paint the dashboard ~60ms after navigation — before DOMContentLoaded and unreliably ordered against init scripts. Both script-based curtain variants lost the race in real recordings (dashboard flash at t=0). The curtain div is injected by route-intercepting the document request and rewriting the HTML.
• route.fulfill() on the document breaks the page's WebSocket: Chromium's Local Network Access checks block ws://127.0.0.1 from a fulfilled (provenance-less) document with ERR_BLOCKED_BY_LOCAL_NETWORK_ACCESS_CHECKS. The recorder launches Chromium with --disable-features=LocalNetworkAccessChecks. Symptom if it regresses: the dashboard renders but stays empty (no tasks, no findings) while API seeding succeeds.

Demo Content Quality

Playbooks

Playbooks in the demo must feel like real workflows a developer would repeat:

• Good: "Implement GitHub Issue" (param: issue URL), "Test Quality Audit" (param: module select), "Security Review" (param: focus area)
• Bad: "Bug Fix" (too generic), "Feature Implementation" (placeholder vibes)

Include at least one playbook with a select dropdown parameter so the viewer sees the parameter form is rich, not just text inputs.

Interactive flows

When demonstrating interactive UI (snooze dialog, launch dialog, playbook parameters):

• Hold the dialog visible for 2-3 seconds before making a selection — the viewer needs to read the options
• After an action changes UI state (snooze moves task to snoozed section, complete moves to completed), wait for the new section to render and hold it visible
• Match production behavior — watch the real app for the flow before scripting the demo. Don't guess at what happens after an action.

Marketing rules baked into the scenario (fresh-eyes review pass)

These came out of a panel of simulated first-time viewers (skeptic, eng manager, newcomer, power user) reviewing the actual frames + narration:

• First 3 seconds show the pain, never a logo. The cold-open grid is on screen from frame 1; the dashboard is revealed with a live finding already queued ("1 ACTIVE") — viewers must never see an empty findings panel.
• The "aha" comes before the tour: permission-block triage is Act 1; project filtering is Act 2. Front-load the hardest problem, not navigation.
• No invented metrics, ever. A footnoted "demo overlay, not a product feature" estimate destroys trust outright. Every quantified claim must be real and on screen (session cost in the TopBar, per-task cost, the completion digest).
• Show the loop closing: after Allow / a sent reply, inject a PreToolUse so the agent visibly resumes. "Send" must not look like it goes into the void.
• Answer "how does it attach?" early (the plumbing narration line: local hooks, replies land in the agent's terminal, nothing leaves your machine). It is the #1 skeptic objection.
• The closing card is a real CTA: repo URL, clone command, star prompt, stat strip. No unverifiable claims ("two minutes") in the narration.
• The thumbnail is captured in a post-credits epilogue after the video_end tracker mark; the published video is trimmed there, so the overlay never appears in footage.

Round 2 (second fresh-eyes panel: skeptic, retention editor, OSS-maintainer ICP, non-native/muted viewer):

• Every spoken claim needs on-screen proof at the moment it is spoken. "One click unblocks" → the reopened agent's terminal streams the granted command running. "This agent is Codex" → the provider badge is opened and zoomed. "CI fails" → the check flips to red live in the zoomed PR card. Claims without visible evidence read as vaporware to skeptics.
• Zoom punches (zoomTo/zoomReset) for anything that must be readable on a phone: permission card + actions, AI reply drafts, PR card, completion digest, the TopBar cost. A 1080p dashboard wide-shot is ~8px text on mobile. zoomTo waits for its selector, so check mode catches zoom-target drift.
• The session cost must visibly tick upward (setSpend at act transitions, final tick lands exactly on the number the narration quotes). A static counter reads as mocked data — the skeptic persona called a never-changing "$1.47" the single biggest credibility hit.
• The cold-open grid must MOVE: pane lines type in progressively, the permission pane pulses. Three identical frames at t=0 loses autoplay viewers before the first caption ends.
• Pulse-highlight (pulseHighlight) interactive elements before clicking them (project chips, GitHub tab) — muted viewers need to see WHERE the interaction happens, narration can't carry it.
• Write narration for B1–B2 English: no idioms ("leverage", "terminal archaeology", "check-in loop"), one verb in the final CTA. ~30% of social-feed viewers watch muted with subtitles.
• Keep the toast on the task that caused it: hold focus ~1.2s after a send before switching selection, or the confirmation toast flashes over an unrelated scene.

Post-Recording Checklist

After recording, verify each of these before committing:

1. Audio: Check [tts] Generated N/N clips in output. If any failed, fix before shipping.
2. Duration: Target 90-150s. Under 60s feels rushed; over 150s loses attention.
3. Images: Open both PNGs and the thumbnail JPG — no broken emoji boxes, blank areas, or wrong state. Thumbnail must be <2MB (YouTube limit).
4. Captions match visuals: Every claim in a caption should be visible on screen at that moment.
5. File size: WebM should be under 20MB at ~150s. If larger, the video may need shorter hold times.
6. Snoozed/completed sections: If the scenario snoozes or completes a task, verify the corresponding section renders in the findings panel.
7. Export matrix: The [export] Artifact summary: block at the end of the run must list every file with no MISSING entries.
8. SRT spot-check: Open kookr-demo.srt — cue times must roughly match when each caption appears in the video.

Publish Checklist

The live video is YouTube DHZrO8T_6Xg, linked from the root README.md (thumbnail image + link). When shipping a new version:

1. Upload to YouTube: kookr-demo.webm (or the 4K MP4), plus kookr-demo.srt as a subtitle track (enables auto-translated captions worldwide), plus kookr-demo-thumbnail.jpg as the custom thumbnail. Title pattern: "Kookr — supervise 5 AI coding agents from one attention queue".
2. Update README.md: replace the YouTube video ID in both the thumbnail img URL and the watch link.
3. Release assets: attach kookr-demo-4k.mp4, kookr-demo.mp4, and kookr-demo-vertical.mp4 to the next GitHub release (per demo/README.md, long-form media lives on releases, not in the repo).
4. Shorts/Reels/TikTok: post kookr-demo-vertical.mp4 with the repo URL in the description.
5. X/LinkedIn: native-upload kookr-demo.mp4 (or the teaser with a link to the full video). Never post a bare YouTube link — native uploads get materially more reach.
6. Update this skill: record the new YouTube ID here and the DEMO_HOOK variant used, so the next regeneration knows what is live.

When to regenerate

• Any visible UI change to surfaces the demo shows (TopBar, findings panel, detail panel, snooze/launch dialogs, GitHub tab)
• A headline feature ships that the narration should mention
• pnpm demo:check fails — fix the scenario, then re-record

Troubleshooting