video-studio
Resolve the requested endpoint before making anything.
Script or storyboard only
Read the supplied material and video-craft, choosing the matching reference: marketing, ecommerce, short film, presenter, music cut, clips, explainer or commentary. Check exact spoken lines with script timing, source facts and the brief. Deliver the requested document. Do not generate media or render for a document-only request.
Finished video
Work through to one finished, postable file. Do not present a plan or preview for approval unless requested, the video exceeds about 2 minutes, or it must follow an exact script; then show the script first. Internal checks and sampled frames still apply.
Find what can be found before asking. Choose length (15–30 seconds for an ad), shape (9:16 for short social video), look, pacing and the user's language when unspecified. Ask only for user-owned missing facts: at most four questions together, with recommended answers. Otherwise proceed and disclose material assumptions.
Material
Choose a short production folder, such as lamp-ad. Import attachments first with ovs timeline import --production lamp-ad --inputs '<JSON array of paths>'; use its local copies and asset ids. Keep assets.json, sourced facts in product.json, media/, and videos/main/{manifest.json,index.html} there. For reference-led work load design-system-importer. Use the host browser to capture real product pages; never fabricate UI or claims.
For every product shot, decide whether existing material actually shows the required action. Reuse suitable footage. Missing product action may need 1–2 generated shots under generation authority. A zoom on a still does not replace required motion. Use product fidelity and image preparation for stills.
After declined, unavailable or exhausted generation, continue only if existing material satisfies the shots or the user accepts a specific alternative. Otherwise explain the missing action, retained work and alternatives before making a substitute. Keep declined spending authority revoked for that request.
Make
Use stage-edit for cuts, joins, reframing, mixing and plain subtitles; stage-decide for selecting what to keep. For designed text, price cards and motion use the timeline and page, with frontend-design for the look. Combine both freely. Put editable words in data-slot elements and keep scene motion independent.
For sound load audio. An ad needs sound unless the user chose silence; use available licensed music or authorized narration. Caption text alone is not recorded speech. Never synthesize music from test tones.
Check and deliver
Run ovs timeline check and ovs timeline snapshot for the current production. Inspect the complete returned frame set against the brief and the relevant craft review; a contact sheet is an index. Fix confirmed problems in a batch before ovs timeline render. Missing or stale visual evidence remains unknown. HyperFrames owns capture and rendering; do not introduce another renderer or imperative media playback.
Run ovs timeline ship --production lamp-ad --video-id main --input lamp-ad/render/main.mp4 for a timeline. For a standalone edited file omit --video-id and supply --expect with durationSec, aspect, audio and optional captions. Inspect frames from the actual final MP4 with ovs edit extract-frame, including opening, every scene, joins and ending. Technical success does not certify visual quality. Ship keeps immutable versions and rejects stale timeline renders.
Repair unmet requirements within available authority, verify the changed output and ship again. If input, authority or runtime prevents completion, report incomplete work, the specific unmet requirement, retained files and what enables continuation. Do not present a broken output as finished or convert findings into an unsolicited enhancement list.
Lead the final reply in the user's language with the video link, then briefly state choices, assumptions, accepted deviations, retained advisories and known provider costs. Keep command names and internal codes out of ordinary delivery prose.
Revise
Change only what was asked at the cheapest level: edit words/cuts/order/music/captions locally; re-voice only a changed spoken line; regenerate only a wrong shot under existing spending authority. Changing caption text does not change spoken audio. Preserve prior delivered versions. Give parallel ratio/language/test variants distinct video ids sharing assets. Reuse current evidence for unchanged bytes, and never repeat an unknown paid request automatically.