The workflow is guided. The agent does the work between checkpoints and recommends an option at each one. You make the decisions. Checkpoints use the ask_user tool. Every answer is appended to branding/decisions-<slug>.json.

Stages

Workspace and brief. video_workspace creates references/, branding/ and outputs/ in the current folder. The agent reads outputs/campaigns/<slug>-{context,persona,psychology,plan}.md when present. It asks a “Shape the brief” checkpoint only if the subject, goal/CTA, length or platform is missing. A stated length is a hard limit; 8 seconds or less becomes a single-beat sting.
Validate the brand. The agent writes branding/brand.html from the skill’s brand template. For a real company it starts from the actual identity, cites its sources in the page, and never invents a logo. You approve it, describe changes, or pick an alternative direction. In the desktop app you can click any element of the page to request a change.
Choose the references. Each new video in references/ is analysed with video_analyze_reference, and the agent writes branding/analysis/<ref>/analysis.md. You pick which references shape the video (multi-select). If references/ is empty, you can add videos, go without references, or ask for reference directions in words.
Pick the style. One checkpoint with up to three questions:
- Style direction: 2–3 directions, each with a preview frame (a video_still from a draft spec, or an image_generate mood frame).
- Generation path: only paths whose engines are connected are offered — motion design (Remotion type and shapes; fastest, free), GPT image + motion (image_generate key visuals with Ken Burns/parallax, about 1 minute per image), or AI footage (Higgsfield video models; slowest, costs credits, with an estimate against the balance).
- Quality: a fast draft render first, or straight to the 1080×1920 master.
Nothing that costs credits runs before this answer.
Approve the storyboard. The agent writes branding/storyboard-<slug>.md and branding/specs/<slug>.json, generates assets into branding/assets/<slug>/, transcribes any voice-over with video_transcribe, and renders hook, middle and last frames with video_still. You render, change something, or try another hook. The master is never rendered before this approval.
Render, finish, QA. video_render, then video_finish. A FAIL verdict is fixed and re-rendered without asking. Each PASS_WITH_WARNINGS warning is fixed or justified.
Review the video. You see outputs/latest.mp4 and choose done, revise, or make a variation (another style, length or platform). Revisions loop back to the stage they touch.

Resuming and non-interactive runs

On resume, the agent reads the decision log and skips checkpoints already decided, unless you ask to revisit one. If nobody is available to answer (a non-interactive run), the agent takes its recommendation and marks that decision inferred in the storyboard.

Subagents cannot talk to you. The lead agent runs every checkpoint itself and only delegates legs that need no decisions, such as reference analysis or the final render, passing the decision log to the video-editor.

Defaults

Format from the brief: face + list → F1, face + thesis → F2, no face + pitch → F3, no face + direct-response copy → F4. Short stings for 8 seconds or less.
Platform all: Reels, TikTok, Shorts and Pinterest encodes.
The workflow never uploads a video or launches ads.

Output

outputs/latest.mp4, outputs/latest-cover.jpg, and outputs/latest.json with the QA verdict
outputs/videos/<slug>/ — master, platform files, QA folder
branding/decisions-<slug>.json — every checkpoint and your answer
branding/storyboard-<slug>.md — ends with a Delivery section: paths, verdict, and decisions taken (inferred ones marked)

See Video Studio for the tools and the QA checks.

Source

Canonical prompt: prompts/video.md.