| name | video-visual-assets-ops |
| description | Design and produce richer visual assets for selected short-form or long-form video beats, including coded motion graphics, diagrams, screen demonstrations, official promo excerpts, generated stills, and generated video. Use when an explicit visual work order needs more explanatory power, novelty, humor, or motion than a screenshot can provide, or when a video asset pack identifies a small number of high-value custom visual opportunities; use the asset-pack skill instead for whole unsorted scripts. |
Video Visual Assets Ops
Turn selected script beats into concise visual treatments and rendered editor assets. Prefer deterministic, editable, source-safe production paths before expensive generation.
Input Gate
Resolve:
- one exact script beat or a small set of selected visual work orders and their intended meaning
- target duration, aspect ratio, and safe margins
- output directory and naming reservation
- whether the visual is evidence, explanation, atmosphere, or comedy
- source material, brand constraints, and available generation tools
If several beats are candidates, rank them by comprehension gain, hook value, production burden, and source risk. Build only the highest-value assets first.
Choose the Production Path
Use production-paths.md to select one of:
- source-faithful screen recording or UI demonstration
- official promo or demo excerpt
- Manim technical explainer
- deterministic Remotion or other coded motion
- diagram or comparison board
- generated still or stylized plate
- generated video
- Blender-authored 3D B-roll or motion graphics
Use screenshots or A-roll when motion adds little. A visual is successful when it clarifies the beat in one glance or creates a deliberate pattern interrupt.
Treat Blender as an explicit opt-in path. Use it only when the creator names
Blender, Blender MCP, or a bespoke 3D treatment in the request. Do not infer
permission from phrases such as make it fancy, cinematic, add B-roll, or
add animation. Prefer generated video for fast cinematic inserts and coded
motion for precise explanatory animation unless the creator deliberately
chooses the additional modeling, lighting, rendering, and QA burden of Blender.
Route mathematical, geometric, equation-based, graph, machine-learning,
algorithmic, and mechanism animations to manim-explainer-ops when semantic
object transformations carry the explanation. Keep Remotion as the default for
UI, brand, typography, screen-like composition, general text motion, and
whole-video assembly. Manim is a specialist asset lane, not a replacement for
the surrounding visual-director or editor workflow.
For a sponsor, brand, company, or product beat, default to a source-faithful
product UI demonstration: a privacy-safe screen recording when access is
available, or the shortest useful excerpt from an official demo or promo
source. Generic coded animation may frame or introduce the demonstration, but
it should not replace showing what the product actually does.
Workflow
1. Write a micro-storyboard
For each selected beat, define:
- first frame
- one visual change or reveal
- final frame
- duration
- narration sync point
- text hierarchy
- provenance or synthetic-media label
Keep most inserted motion between 1.5 and 5 seconds. Avoid miniature explainer films inside a short unless the visual is the premise.
For talking-head footage, keep proof and single-state motion partial-frame by
default. A full-screen asset must have an explicit legibility or comprehension
job and a meaningful change between its first, middle, and final states; a
static card or one looping scale move does not qualify.
Inherit the accepted title composition when it defines the video's visual
stage. If the title sits center-bottom and that region is face-safe, design the
asset's focal action for center-bottom and keep it large enough to read at
phone playback size. Do not default to a side card merely because the asset is
partial-frame. Validate one real A-roll composite before rendering a family of
assets.
2. Fan out when paths are independent
Use available subagents for bounded lanes such as:
- visual-concept alternatives
- official source/promo inventory
- coded animation implementation
- independent legibility and motion QA
Give each worker a reserved filename or staging directory. Do not have multiple workers edit the same source file. The parent selects the concept, merges provenance, and performs final visual inspection.
3. Prefer reversible local prototypes
Without additional approval, create lightweight local drafts using available code, SVG/HTML/CSS/canvas, ffmpeg, screen capture, or existing generation entitlements when the user already asked for asset production.
Ask before:
- billable image or video generation
- long generated-video jobs
- large source-media downloads
- generation that materially changes the factual framing
Do not pause for approval merely to make a small local poster frame, motion test, or storyboard.
Choose the production engine by visual job. Use source-faithful screenshots,
screen recordings, or official demo footage when the asset proves a product or
claim. For cinematic inserts, atmosphere, transformation shots, or clearly
illustrative metaphor, prefer the authenticated Gemini consumer video surface
when generative video materially improves the shot. Use Remotion or another
coded-motion path for UI, brand, text-led motion, routing diagrams, readable
labels, exact narration sync, and reusable brand treatments. Use
manim-explainer-ops for precise mathematical, geometric, graph, algorithmic,
equation, machine-learning, or mechanism animation. Keep Higgsfield parked
unless the user separately approves its paid subscription for a specific shot.
Generated footage must not impersonate product proof or a real event.
Use Blender MCP only after the explicit Blender/3D gate above passes. Reserve
it for shots whose value comes from controllable 3D geometry, camera moves,
lighting, depth, simulation, or reusable scene assets. Start with a short
storyboard and low-resolution proof render, keep the Blender source file, and
do not let scene construction block the usable first edit. Treat Blender
output as synthetic illustration, not documentary or product evidence.
A paid consumer Gemini plan does not imply that Gemini API, Vertex, MCP, or a
third-party generator draws from the same allowance. For subscription-backed
Omni/Veo generation, default to the signed-in Gemini Apps browser surface at
gemini.google.com and use its Videos workflow. Do not route consumer video
through the legacy Gemini CLI: Google stopped serving consumer, Google AI Pro,
and Google AI Ultra logins there on 2026-06-18. Antigravity CLI is the supported
replacement for coding-agent work, not a documented headless Omni/Veo surface.
Before submitting a Gemini Apps video run:
- use the active browser-control skill, never desktop Computer Use merely to
operate Gemini Apps
- verify the intended Google account and visible Google AI plan
- inspect
Settings > Usage limits when available
- use the
Videos / Create with Omni surface
- confirm that no API-key, Vertex, third-party credit, purchase, upgrade, or
paid-overage path is being invoked
Browser-based Gemini Apps generation consumes the plan's product quota and is
not a per-second Gemini API call. If the browser surface is unavailable or the
plan quota is exhausted, preserve the prompt and ask before switching to API
billing, purchased AI credits, or another paid provider. Record provider,
displayed model, prompt, signed-in account class (not private credentials),
billing/quota surface, and synthetic status in the asset manifest.
Treat generation limits as a visible production budget. Before a billable or
quota-consuming run, tell the user which engine will be used and why. After the
run, report the assets produced, attempts consumed when observable, and the
remaining allowance when the product exposes it. If the allowance is hidden,
say so instead of estimating. Prefer a small number of well-scoped candidates
over repeated speculative generations.
For an opening-hook candidate, require a readable first frame and make the
shortest editor derivative that performs the job—usually about 1–3 seconds,
even when the model returns a longer clip. Inspect the entire accepted range
for morphing objects, garbled text or logos, provider marks, waxy materials,
and glossy sci-fi styling that makes the asset feel generically AI-generated.
Trim, crop, or reject those ranges before CapCut integration; do not keep a
longer generation merely to use more of the paid output.
4. Preserve truth and provenance
- Never generate a fake receipt, article, benchmark, product UI, or social post.
- Do not present generated footage as a real event, person, product demonstration, or historical record.
- Keep synthetic visuals clearly separated from evidence assets in the manifest.
- Prefer official press kits, product pages, or creator-owned source footage for promo excerpts.
- Record source URL, owner, downloaded file, excerpt in/out points, and any visible watermark or usage limitation.
- Use the shortest excerpt that performs the editorial job.
5. Compose existing skills
- Use
manim-explainer-ops for precise mathematical, geometric, graph, equation, machine-learning, algorithmic, and mechanism animation.
- Use
social-motion-diagram-ops for sketch-style diagrams and small looping flows.
- Use
social-video-archive-ops when official or reference video must be downloaded and preserved; create only the short editor-ready derivative here and retain exact in/out points.
- Use an image-generation skill for generated stills or image edits when available.
- Use the active browser or screen-control skill for real UI demonstrations.
- Return to
video-asset-pack-ops so the rendered files and provenance enter the numbered manifest.
6. Render editor-ready deliverables
Prefer:
- PNG poster or first frame
- MP4 for motion and generated video
- editable SVG, HTML/CSS, or source project when practical
- alpha-capable formats only when the editor workflow supports them
Use stable names such as:
NN_<beat>_poster.png
NN_<beat>_motion.mp4
NN_<beat>_source.svg
NN_<beat>_promo_excerpt.mp4
7. Verify
Inspect the first, middle, and final frames. Probe the rendered file with the
video-asset-pack-ops scripts/inspect_asset_pack.py helper and preserve the
reported codec, dimensions, duration, frame rate, audio streams, and rotation.
Verify file size, readable text, safe margins, clean loops, and absence of
blank frames. Confirm that the visual still makes sense with the sound off.
Output Contract
Return:
- ranked visual concepts and chosen production path
- rendered asset paths
- editable source paths when available
- duration, dimensions, and verification results
- provenance or synthetic-media status
- exact timeline placement and narration sync point
- approval-required concepts that were not run
Guardrails
- Do not build motion solely to make the edit busier.
- Do not let custom visuals block the usable first edit.
- Do not imitate a living creator's proprietary style exactly; extract general visual mechanics.
- Do not conceal missing evidence with generated media.
- Do not invoke paid or scarce generation silently.
- Do not invoke or propose Blender merely because a richer visual might look
impressive; require an explicit Blender, Blender MCP, or bespoke 3D request.
- Do not copy private media into public repositories.
- Do not own project folders, raw-media intake, narration lock, or whole-video coverage decisions.