Build or edit a video with a Director that can assemble exact source segments, inspect rendered frames at scene holds and cuts, report specific visual faults, and verify one bounded repair cycle instead of judging only timeline metadata.
Where: Video project → AI assistant → ask to build a video or cut/edit connected source footage
Key ideas
- Source footage becomes an edit, not just a background: The footage stage can analyze a connected video, choose source in/out ranges, assemble them back-to-back with correct master placement, reframe a drifting subject for the output aspect, and apply an editable speed ramp.
- The art director sees rendered evidence: The build captures each scene’s mid-hold plus frames just before and after cuts from the same bundle substrate used by rendering, including fonts and decoded video frames.
- Findings use a closed vocabulary: The visual pass can report unreadable text, cropped subjects, collisions, dead-center drift, flat backgrounds, type that is too small, missing media, and empty frames. Unknown free-form labels cannot trigger repair actions.
- Repair follows real levers: Missing media can be refilled; empty frames and unreadable text can enter the bounded timeline/CSS repair path. Composition judgments stay warnings when the current build has no safe automatic redesign.
- Verification is explicit: The Director re-runs deterministic checks after repair. CSS legibility patches are captured again and reverted when the finding persists, so a changed stylesheet is not treated as proof by itself.
Steps
- Start a Video project and attach the brief, brand/reference material, and any source footage that should be cut rather than merely used as inspiration.
- Ask the Video Director to build the piece or cut/edit the source; state the delivery aspect, target duration, audience, and what must stay visible/readable.
- Review the script/cut plan and confirm which source ranges are kept before expensive generation or detailed design.
- Let the build design scenes, assemble footage, apply motion/transitions, fill media, and add audio/captions.
- Inspect the progress/notes for frame-review availability, coded visual findings, and the single bounded repair cycle.
- Use Verify Render at representative scene holds and cut points when manually refining the result; request a captured frame when visual proof is useful.
- Manually fix warnings that require editorial/design judgment, then review the actual rendered output from start to finish.
Source-footage tools
- Analyze Video: timestamped scene/visual/audio description.
- Assemble Segments: exact source in/out ranges placed gaplessly on the master timeline.
- Reframe Clip: editable crop/translate/scale tracking for a subject that would drift out of frame.
- Set Clip Speed Ramp: time-remap keyframes while preserving timeline duration.
- Plan Cuts: global transition and camera-rhythm pass.
- Verify Render: scrub, measure real on-canvas geometry, report issues, and optionally capture a frame.
Visual finding codes
- text-illegible — blocking; text cannot be read at a glance.
- media-hole — blocking; expected media is blank/broken/placeholder.
- frame-empty — blocking; a hold moment has essentially no readable content.
- subject-cropped — warning.
- composition-collision — warning.
- dead-center-drift — warning when the ordered composition asked for another structure.
- bg-flat — warning when the ordered background lacks its intended treatment.
- type-too-small — warning.
Evidence plan
Mid-scene frames are load-bearing because they show the intended hold. Pre- and post-cut frames catch transition gaps, jumps, and near-identical adjacent shots. When the frame cap is reached, mid-scene coverage is preserved before later cut enrichment.
Repair boundaries
The pipeline runs at most one repair cycle. A code becomes blocking only when a concrete lever can act on it; otherwise it remains a warning for human/editorial correction. If the visual model or capture path is unavailable, the build records that fact and skips vision-dependent correction rather than claiming a visual review occurred.
Render and repair reliability
Recent fixes address composition finishing, attached precomp identity, image-sequence readiness/CORS, and initial frame rendering. Embedded document editors retain access to the render project. Video repair targets existing scenes and source reads instead of unnecessarily recreating shots or regenerating already available media.
- A dropped frame folder creates an image sequence with playback configuration rather than only a static frame.
- Wait for image sequences to settle before judging captured output; the render path now handles their readiness explicitly.
- Check the actual browser/desktop/GPU result after changing nested compositions; a source preview alone does not validate delivery.
- Caption work is bounded, and the editor reduces redundant generation and idle timeline rebuilding.
Tips
- Name which source moments are mandatory or forbidden before the cut plan.
- Use a target aspect early so subject reframing and shot composition are judged against the real delivery frame.
- Treat the final verifier as a floor, not final creative approval: watch the rendered video with audio at the target size.
- Check opening hook, every cut, CTA legibility, captions, and the first/last frames manually.
Limitations and important notes
- The art director has a finite frame budget and does not inspect every frame of a long video.
- Only one bounded automatic repair cycle runs.
- Some composition warnings require a human redesign because no safe automatic lever exists.
- If DOM capture, the bundle, frame capture, or the visual model is unavailable, pixel-aware repair is skipped and the notes explain why.
- Auto-reframe can leave a clip unchanged when it cannot find a trackable subject.
Troubleshooting
The build says the vision gate was unavailable
Review the named reason in the notes, verify the render bundle/frame capture works, and perform manual Verify Render checks before delivery.
Unreadable text was changed but still fails
The CSS patch should revert when confirmation still finds the problem. Manually add contrast/scrim, reposition the text, enlarge it, or simplify the background, then verify again.
Source clips land at the wrong master times
Use Assemble Segments rather than placing each trimmed clip manually; it owns the source-in and master-placement math.
A vertical crop loses the speaker/product
Run Reframe Clip after placement and review the editable tracking/crop keyframes across the whole clip.
The verifier passes but the video still feels weak
Passing means the deterministic/blocking checks cleared. Address composition and craft warnings, pacing, story, brand fit, and editorial judgment manually.
FAQ
Does the Video Director inspect the editor canvas or the actual render?
Its frame review uses a sandboxed render bundle aligned with the export substrate, including real fonts and video frames, rather than relying only on editor metadata or placeholder boxes.
Why are cropped subjects and composition collisions warnings?
The current bounded repair has no safe automatic redesign lever for them. Marking them as blocking would fail a build without giving it a reliable way to improve.
Can I ask for a frame during verification?
Yes. Verify Render can capture the current timecode as visual proof in addition to returning geometry and issue data.