Vox Director: One-Topic Vox-Style Collage Videos
Turn a single topic into a finished Vox-style paper-collage explainer or ad video — script, keyframes, motion, voice-over, music and captions, all automated.
SKILL.md stops when ATLASCLOUD_API_KEY is missing, enforces two human approval gates (beat map, aspect approximation), minimal local deps (ffmpeg/Pillow), no credential exfiltration signs; but it uploads user photos, talking-head video and voice-clone samples to the third-party Atlas Cloud API, and the default watermark is 'Made with Atlas Cloud' (promotional), with incomplete data-flow disclosure — deducted.
Extensive documented gotchas, auto-resubmit, model fallbacks, resumable generation (skips existing keyframe_url) show serious engineering; but no committed tests or CI were seen, key paths are unverifiable statically, and error handling exists only as prose — capped at 10 for static review.
Triggers, three input modalities (B/A/C-roll) and boundaries (Kling for real people, content-filter workarounds) are very clearly written, bilingual; but the core function depends entirely on the single overseas Atlas Cloud API (paid, quota-bound) with no mainland-China reachability statement — deducted.
Clear layered architecture (SKILL.md + references), MIT license, package. version, bilingual docs; but no changelog, maintenance is by one individual, update path unstated, and hidden assumptions (Atlas account, drifting model IDs) are delegated to external docs — deducted.
If executed as documented, it produces a finished video end to end with clear marginal value (manual collage video production is far costlier); but no independent reproduction statically, and cost/quality evidence is author-supplied showcase only — capped at 6.
Example beats., showcase videos and concrete recorded failures (e.g. 'PARFUM'→'PAREUM') exceed pure marketing; but everything is author-attested, with no third-party tests, CI, or independently checkable model verification — capped at 4.
- Core function depends entirely on the paid Atlas Cloud API; mainland-China reachability and account availability are undeclared — verify before use.
- User photos, talking-head video and voice-clone samples are uploaded to a third-party cloud; biometric/voice data implications should be assessed by privacy-sensitive users.
- Default output carries a 'Made with Atlas Cloud' watermark (promotional); confirm removal for commercial use.
- Model IDs drift; docs admit defaults work 'today' only, so manual model reselection may be needed at runtime.
- Voice cloning and the 'zero content-filter' local engine path for animating real people carry misuse risk (deepfake/impersonation); users must ensure compliance themselves.
What does this skill do, and when should you use it?
Vox Director is an agent skill that turns a one-line topic into a finished Vox-style torn-paper collage explainer or ad video, end to end on the Atlas Cloud API plus local ffmpeg. The pipeline is driven by one beats. per project: pick a narrative arc, run a style bake-off, generate collage posters, animate them, add narration and music, and assemble with ffmpeg. Beyond B-roll (topic to video), it supports A-roll (re-styling an existing talking-head video into the collage look) and C-roll (anchoring a single photo of a person or product as a photographic sticker inside the collage). Two human decision gates keep you in control: approving the beat map and picking the visual style by eye.
Each stage is one script: 1) writes beats. using narrative arcs from beat-layer.md (timeline, pas, how_it_works, etc.) and shows it for approval; 2) style_bakeoff.py renders the same beat in 3–4 theme presets for a visual pick; 3) keyframes.py calls google/nano-banana-2/text-to-image to produce a collage poster with baked-in headline per shot (about $0.08 each, cheap to re-roll); 4) clips.py animates each poster with gemini-omni-flash image-to-video, applying per-shot camera moves and rich element_motion (real people/brands are routed to kling-video-o3-pro); 5) audio.py generates narration via xai/tts-v1 (optional seed-audio voice cloning) and instrumental BGM via minimax/music-2.6; 6) assemble.py uses ffmpeg to concat, duck music under VO, and burn captions plus a watermark into final.mp4. All calls go through scripts/provider.py, which auto-resubmits stalled or failed jobs.
- A creator who already has a talking-head video and wants it re-styled into the collage look while keeping the real face, lip-sync and gestures (A-roll)
- A brand or indie marketer turning one product photo into a collage ad, with the label typography kept pixel-faithful (C-roll)
- A knowledge/explainer channel needing dense 9:16 vertical social videos with a cut every 4–6 seconds
- Users not on Atlas Cloud who want a manual prompt-pack: beat map + per-beat image prompts + motion prompts + narration script to paste into any generator
- History/tech/food topics needing multilingual narration (xai/tts offers 5 multilingual plus ~66 native voices)
- When you need dramatic piece-by-piece assembly motion or filter-free animation of real people via the local keyframe engine
What are this skill's strengths and limitations?
- Fully automated end to end — one topic in, a finished mp4 out — while keeping two human checkpoints (beat map, style pick)
- One script per stage driven by a single beats.; after approval, re-runs only generate what's missing (keyframes.py skips shots that already have a keyframe_url)
- Exceptionally thorough docs: prompt structures, 14 narrative arcs, a voice roster, and a catalog of every API/ffmpeg gotcha
- Pluggable media backend (provider.py) — switching backends requires no changes to stage scripts; stalled jobs auto-resubmit
- Cost-aware design: re-roll weak keyframes cheaply ($0.08) instead of paying to animate a bad image
- Hard dependency on Atlas Cloud — the only backend today; without its API key you're limited to the manual prompt-pack mode
- SKILL.md gives no overall cost estimate for a full film and mentions no automated test suite
- Content restrictions exist: Omni and Seedance refuse real people and brand logos (use Kling instead); in A-roll you must never ask the model to redraw the face
- Model IDs drift, so the live model list must be fetched before each run
- Some behaviors are known traps (e.g. seed-audio cloning requires the pinned-speaker template) and require reading the gotchas doc to avoid
How do you install this skill?
Prerequisites: any coding agent that can follow a workflow and run scripts (Claude Code, Codex, etc.), an Atlas Cloud API key, ffmpeg + ffprobe, and Python 3 with Pillow. Option A: git clone https://github.com/Alisa0808/vox-director.git ~/.claude/skills/vox-director. Option B: download vox-director.skill and install it via your Claude skills UI. Then export your key: export ATLASCLOUD_API_KEY="sk-..." (get one at atlascloud.ai/console/api-keys). Non-Claude agents read AGENTS.md → SKILL.md.
How do you use this skill?
With the skill installed, just ask your coding agent, e.g.: "Make me a Vox-style collage video introducing Mexican street food — English, 16:9, 15 seconds." The agent drafts a beat map for your approval (gate 1), runs a style bake-off for you to pick by eye (gate 2), then generates keyframes → motion → voice → music and outputs out/<project>/final.mp4. To verify the result, extract frames to jpg with ffmpeg — you cannot read an mp4 directly.
How does this skill compare with similar options?
The README explicitly positions the look and workflow as inspired by Vox's explainer visual language and the collage-ad workflows of Stav Zilber, rom1trs and Higgsfield — the skill automates those workflows rather than competing with a named alternative. The source lists no other comparable tools.