Devrel Video

DevRel Video Studio — /devrel-video

A thin orchestrator. It does not reimplement capture or composition — it drives:

• playwright-cli (page.screencast) for live interaction capture,
• HyperFrames for composition + deterministic MP4 render,
• FFmpeg / ffprobe for convert, master, and QA.

Read references/production-notes.md first for environment gotchas (sandbox flags, init flags, CDN vendoring, determinism) and the production steps for music, narration, avatar, subtitles, and multi-environment (terminal ⇄ browser) (all supported — steps 5.4–5.8; the template and manifest carry seams for each).

For studio installation and prerequisites, see the repository-root docs/installation.md. The quick repository demo in examples/README.md demonstrates intake only; use the production workflow below for actual videos.

When to use

• "record / make / produce a walkthrough, demo, catalog tour, or promo of <web product / URL>"
• "update / re-render an existing devrel-video project"

Inputs — gather, never guess

• URL (+ start command if the product is local).
• Flow: the 2–4 beats to show (e.g. browse → filter → open a detail).
• Brand: reuse brands/<name>/, or extract from the URL (step 3).
• Output: format(s) + duration target. Default: silent 16:9, ~25–40 s.
• Environments (optional) — a demo can switch between the browser and Claude Code (terminal). Name the order in the prompt ("prompt in Claude Code → the catalog in the browser → back to Claude Code") → environments: in brief.yaml. Terminal beats are captured with VHS (step 5.8); a pure-browser video is the default.
• Audio / avatar (optional, all prompt-selectable) — set any of these straight from the prompt, and write them into brief.yaml:
  • voice: the narration voice — male leo/dan/zac, female leah/tara/jess/mia/zoe (match the presenter; offer samples if unsure). "…narrated with voice dan" → narration: dan. Regenerating with a different voice makes a new variant (see 5.4).
  • music on/off, subtitles languages, avatar at intro/CTA (+ its private avatar_plate path).
• Approved claims: numbers/names allowed on screen — pull from the captured page, never invent (guide §7/§8).

Workflow

1 · Intake → projects/<slug>/brief.yaml

Copy templates/brief.example.yaml. Confirm the audience, the one message, the CTA, and the beats.

2 · Explore (before recording)

Open the URL headless; snapshot + screenshot. Confirm: no login wall, no PII / keys / wallets on screen, and stable selectors for each beat. Prefer exact a[href="…"] and getByRole locators; always add a fallback selector so one missing element can't abort the capture.

3 · Brand → brands/<name>/

scripts/brand-extract.sh <url> <name> → tokens.json (real palette + fonts) + logo.svg. Use these in the composition. Never invent a palette. Reuse an existing brands/<name>/ if you already have one for the product (brands/acme/ is the default placeholder brand).

4 · Capture → projects/<slug>/capture.mp4

Copy templates/hero-script.js → capture.js; edit the beats (eased scrollTo, real locators, deliberate pauses, all-forward navigation so page.screencast survives route changes). Then: scripts/capture.sh projects/<slug>/capture.js "<start-url>" → records capture.webm, converts to capture.mp4. Contact-sheet it; re-record on layout shift, PII, or a mis-click.

5 · Compose → projects/<slug>/index.html

Copy templates/composition.html to projects/<slug>/index.html and fill it from the brief + tokens (skip hyperframes init: it refuses a folder that already holds the capture, and the render only needs index.html): title card (real claims) → footage in a browser frame → chapter lower-thirds synced to the capture beats (+ the title-card offset) → CTA. Give every animated element a stable id (GSAP and lint both require it).

5.4 · Narration (optional — drives the timeline)

For a narrated cut the voice sets the timing. Write ~1 line per beat (clean text: no parens/version numbers, em-dash → comma). Synth each with Orpheus in a voice that matches the presenter — male leo/dan/zac, female leah/tara/jess/mia/zoe (offer samples if unsure): scripts/tts.sh "<line>" vo1.wav <voice> (set ORPHEUS_URL/ORPHEUS_KEY — see scripts/_env.sh; reference TTS is orpheus-tts). Then adapt projects/<slug>/narrate.py (copy the template — footage length, beat times, SPOKEN lines, captions) — it measures the lines, computes scene/chapter times, and writes voice.wav, re-timed captions.<lang>.srt, and timing.json. Apply that timing to the composition with scripts/retime.py projects/<slug> (idempotent — rewrites every data-start/data-duration + GSAP position by id, so total tracks the voice), then mix (next step), render. Swapping the voice later is one command: scripts/variant.sh projects/<slug> <voice> runs the whole chain — re-TTS → narrate → retime → mix → re-lip-sync the avatar bubbles (if any) → render → renders/out-<voice>.mp4 (add --no-avatar to skip the GPU step). narrated-devrel mode.

5.5 · Music (optional)

Local, license-clean bed — no auth, no model: python3 .claude/skills/devrel-video/scripts/gen-bgm.py bed.wav && ffmpeg -i bed.wav -b:a 192k projects/<slug>/bgm.mp3, then add the <audio id="bgm" class="clip" src="bgm.mp3" … data-track-index="9" data-volume="0.85"> seam to the composition (lower the volume under narration). For a produced track: /media-use resolve --type bgm (HeyGen catalog — needs heygen sign-in) or --local-only (MusicGen), dropped into the same seam. With narration, don't stack two audio clips — pre-mix: scripts/mix-audio.sh bgm.wav voice.wav narration-mix.mp3 ducks the bed under the voice and masters to −16 LUFS; use that one file in the seam at data-volume="1".

5.6 · Subtitles (optional)

Author caption cues timed to the beats → projects/<slug>/captions.<lang>.srt — soft, toggleable, platform-standard; one file per language covers "alternative subtitles". Ship the .srt next to the video (YouTube/LinkedIn read it). For autoplay-muted social, burn a hard cut: scripts/subtitles.sh projects/<slug> captions.en.srt. When narration exists, derive the cues from its transcript (WhisperX via /media-use) instead of authoring.

5.7 · Avatar (optional — face-only presenter)

A face-only circular bubble, lip-synced, shown ONLY at intro / one transition / CTA (product stays the hero — a full-time presenter reads as a webinar). Depends on narration. Trim the exact line the mouth should say from narration-mix.mp3 so lips match what's heard: ffmpeg -ss <t0> -to <t1> -i projects/<slug>/narration-mix.mp3 -ar 16000 -ac 1 line.wav. Then scripts/avatar.sh <plate> line.wav projects/<slug>/avatar-face.mp4 — runs LatentSync (256) on your own GPU host over SSH (set AVATAR_HOST/LS_DIR — see scripts/_env.sh; reference is an AMD ROCm box), lip-syncs, then pins the face (per-frame face-tracking via track_crop.py — follows the face every frame at constant size, so the head stays put in the small circle instead of drifting; vidstab alone can't fix a moving subject). Add a <video class="clip face-bubble" src="…" data-start data-duration data-track-index="11"> element with a shared .face-bubble CSS class (border-radius:50% = the circle) + a GSAP pop-in / fade-out over that window (data-duration = the line's length). For a bubble at more than one moment (e.g. intro and CTA), run avatar.sh once per line and add one element per window; set .face-bubble z-index above any full-screen scene card it overlaps (the CTA card is z-index 8 → bubble z-index 9). The <plate> is the avatar_plate: from brief.yaml — a private recording kept OUTSIDE the (public) repo (convention: ~/Videos/<Org>/DevRel/Talking_Head/<name>/); generated avatar-*.mp4 bubbles are gitignored, never committed. Format: front-on, sitting still, even light, mouth visible, ≥ the line's length; any fps (avatar.sh normalizes to 25 fps — LatentSync's rate, else the lips drift). Brighten a backlit plate first (ffmpeg -vf eq=brightness=0.06:contrast=1.10:gamma=1.12), and raise STEPS=30 (env, default 20) for a crisper mouth. ~10 min/clip cold. narrated-devrel + avatar.

5.8 · Multi-environment (optional — terminal ⇄ browser)

For a demo that moves between the CLI and the web (e.g. prompt in Claude Code → the product in the browser → back to Claude Code): capture each terminal beat with scripts/capture-terminal.sh <tape> — VHS renders the real Claude Code TUI headlessly (pixel-perfect; no sandbox flag needed) to term<N>.mp4. Copy templates/terminal-script.tape; set the demo dir + prompt. Sanitize like step 2 — the terminal is the CLI's screen: record in a clean throwaway dir (host paths show in the prompt and in tool calls), unset CLAUDE_CODE_CHILD_SESSION (kills the "transcript saving is off" warning), alias claude to claude --strict-mcp-config (hides the "N MCP servers need authentication" line — the flag lives in the alias so the recorded command stays a clean claude), and Escape onboarding popups. A live model turn is non-deterministic (wording + surprise popups vary) — for a repeatable cut, script a fixed answer (the HTML-terminal re-enactment — templates/cc-terminal.html) or record a good take and trim. Capture browser beats as usual (step 4). Then compose: place the beats as sequential timeline clips with terminal-frame / browser-frame chrome and hand off by fading each beat through the brand teal (the .bg shows through) — a pure crossfade ghosts terminal text through the page; fading over .bg is clean and adds the scale-pop the browser cut uses. See the MULTI-ENVIRONMENT seam in templates/composition.html. Quick non-composition preview: scripts/stitch-envs.sh out.mp4 0b1020 term1.mp4 capture.mp4 term2.mp4. Worked example of the re-enactment approach: templates/cc-terminal.html (a self-contained, renderable Claude Code beat). Tooling: VHS (go install) + ttyd (auto-fetched by the script) — for a free-form unscripted real session, record with asciinema and render with agg instead.

6 · Render + QA → projects/<slug>/renders/

scripts/render-qa.sh projects/<slug> → high master + web-compressed copy + ffprobe + contact sheet. Gates before "done":

• hyperframes lint 0 errors · hyperframes check passes (Contrast AA).
• ffprobe: target dims/fps/duration, video stream present, non-zero size.
• Contact sheet: no clipped text, correct product state, chapters aligned, no PII.
• Security: no keys/wallets/PII; all CDN assets vendored for confidential projects.
• Terminal beats: sanitized — no host paths, no "transcript saving is off" / MCP-auth warnings, no onboarding popups on screen.

6.1 · Thumbnail (required deliverable)

Every video also ships a YouTube thumbnail (not just the QA contact sheet): copy templates/thumbnail.html → projects/<slug>/thumbnail.html, fill it from the brief + brands/<name>/tokens.json (kicker, headline, url label), then scripts/thumbnail.sh projects/<slug> → renders/thumbnail.jpg (1280×720, <2MB; pulls the hero shot from capture.mp4). For a README/social loop, a short GIF off the master: ffmpeg -ss <t> -to <t> -i renders/out.mp4 -vf "fps=12,scale=880:-1,palettegen" … → paletteuse.

Delegation map — do not reinvent

Roadmap

Music, narration, avatar, subtitles, and multi-environment (terminal ⇄ browser) are all supported (steps 5.4–5.8). Still open: portrait/social output profiles (9×16 recompose, not crop) and a short muted web-loop. See references/production-notes.md.