google-omni-video
Skill0 starsChanged 13 days agoMITView on GitHub
Design and generate reference-led promotional shots with Google Omni Video (Gemini video models), preserving real brand and product assets. Use for Omni video briefs, native-audio generation, visual continuity, shot revisions and multi-shot film assembly; not HTML animation or local video editing.
In this repository34 other files it gives its agents.
- chatgpt/AGENTS.mdAGENTS.md
- opencode/AGENTS.mdAGENTS.md
- zcode/AGENTS.mdAGENTS.md
- claude/CLAUDE.mdCLAUDE.md
- agent-doc-authoringSkill
- agent-experienceSkill
- ai-image-promptsSkill
- api-behavior-flowSkill
- architecture-and-performanceSkill
- behavioral-testingSkill
- behavior-auditSkill
- browser-tabs-uiSkill
- code-qualitySkill
- context-compactSkill
- design-principlesSkill
- design-to-deliverySkill
- design-validationSkill
- e2e-test-authoringSkill
- engineering-reviewSkill
- excalidrawSkill
- external-ai-consultSkill
- llm-agent-tool-disciplineSkill
- parallel-subagentsSkill
- qa-test-scenariosSkill
- reliable-executionSkill
- research-and-reuseSkill
- server-security-hardeningSkill
- session-handoffSkill
- shadcn-librarySkill
- sudo-handoffSkill
- task-intakeSkill
- tauri-v2-docsSkill
- test-failure-triageSkill
- user-behavior-flowSkill
The file
We split this skill into the rules it gives an agent
10 rules. About 0 are common, 10 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈10
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- could be automatic: a setting could enforce it
- Distinctive: line 9
Create expressive video shots with real visual identity, controlled paid iteration and inspectable results.
- Distinctive: line 12
Use for Google Omni video generation and revision, and for assembling multiple generated shots into one film.
- Distinctive: line 15
Read the current user brief, originals, approved adjacent shots and project rules.
- Distinctive: line 32
Write the shot brief: reference roles.
- Distinctive: line 42
Prefer a fuller scenario line over a short label line when the shot must fill10s.
- Distinctive: line 45
If two pacing retakes fail, change the dialogue length/structure, not just the adverbs.
- Distinctive: line 50
After every reassembly, verify master sources by hash, not by file presence: a rebuild can silently use a superseded clip for an …
- Distinctive: line 56
Test contrast for each logo, including all its component colors.
- Distinctive: line 57
Do not hardcode any one brief's palette, architecture, shot length or card layout as a universal style.
- Distinctive: line 68
Check the prepared input first, then actual output.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Google Omni Video
Purpose
Create expressive video shots with real visual identity, controlled paid iteration and inspectable results. Treat references as design inputs, not evidence that the model preserved them.
When to use
Use for Google Omni video generation and revision, and for assembling multiple generated shots into one film. General image briefs can use ai-image-prompts as a companion; this skill owns Omni execution and its video-specific decisions.
Preconditions
Read the current user brief, originals, approved adjacent shots and project rules. Identify protected content, allowed changes and the authorized generation scope. Verify current official Google documentation and account model access before relying on stored model names, prices, schemas or limits.
Versioning strategy — explore cheap, finish real
At session start, ask the user one routing question (skip only if the brief already answers it): "Creative-first or reference-first?"
- Creative-first → build v0.1/v0.2 as unconstrained creative probes (no real project components), then v0.3 as the bridge with real assets, then the final. Use when the visual direction is open.
- Reference-first → use the real references from the very first take; the ladder compresses to (exploration still allowed but anchored to real assets) → final. Use when identity/UI fidelity matters from the start or the direction is already known.
Recommend this ladder to the user when a brief is open-ended: exploratory versions are cheap decision tools, the final is the only expensive one. Never present an exploration as shippable, and never spend final-round care (hash-verified sources, transcription gates) on an exploration.
The ladder itself:
- v0.1 / v0.2 — creative explorations. Deliberately unconstrained style probes (different visual directions, no obligation to real product geometry). Their job is to reveal what the story wants; expect to discard both.
- v0.3 — the bridge. Same creative energy, but now built from real project components (actual UI crops, actual mascot/brand assets in the prepared compositions) — still experimental in style.
- Final (v1.0) — the realistic build. Clean-base method: approved base image + prepared per-shot first/last compositions holding the real assets, one native clip per shot, joined by short dissolves. This is the decision-ready version.
Workflow
- Separate identity, product evidence, environment/style and composition/timing references. Inspect their actual colors, transparency and visible text.
- Choose a visual idea fitting the story. Use crops when only a small UI detail matters. For exact logos or lettering, prepare a finished input composition using originals rather than requesting model redraws. Inspect it before paying for video.
- Write the shot brief: reference roles; protected regions; allowed motion; timed visual beats; framing; spoken line and sound; exclusions. Use first/last-frame controls when supported. Keep impossible precision promises out of the brief and reporting.
- Carry continuity through shared materials, light, palette, motion and audio. Design neighboring boundary frames deliberately; verify the generated transition rather than assuming continuity from prompt text.
- Generate only within authorized scope. Persist submission intent before POST, then operation ID and result metadata before download. Resume polling/downloads without resubmitting. Stop on uncertain submission state or an exhausted allowance.
- Inspect representative frames and moments most likely to fail; compare originals, lettering, contrast, UI and mascot identity. Verify speech by measurement, not listening assumptions: run
ffmpeg silencedetect(e.g.-30dB:d=0.8) for dead air and transcribe with a speech model (e.g.gemini-3.5-transcribe:generateContentwith inline WAV) to confirm the exact words, order and absence of stutters. Frame sampling alone does not cover the audio half of a native-audio shot. - Deliver the native output with a bounded verification statement and approval state. Preserve useful prompts and lessons. Apply the user's retention policy: keep selected outputs, consolidate references/history, and remove rejected or superseded media only when authorized. Avoid a new top-level folder for every prompt draft.
Native-audio pacing (earned over multiple paid takes)
Short dialogue invites dead air: the model speaks a short greeting then idles in silence, and the pad can drop out with it — users hear this as "dry/robotic" even when frames look perfect. Fix in the brief, then verify by measurement:
- Demand speech spanning almost the whole clip (e.g. "speech from0.3s THROUGH9.6s of a10s shot"), phrase breaks shorter than0.3s, and the rest pose only in the last0.3s.
- Attach the opening word to the following clause ("Welcome." directly into the sentence); an isolated greeting is the recurring dead-gap point (observed 2.63s, then 1.46s, gone with the longer connected script).
- Prefer a fuller scenario line over a short label line when the shot must fill10s; ask the user for the scenario when pacing keeps failing.
- Require the ambient pad to start at0.0s and sustain unbroken to the end, never dropping out between sentences.
- Copy the proven pacing block verbatim across sibling shots so sound design matches; verify each take with silencedetect + per-second RMS, and transcribe (a stutter in the transcript is a robotic-delivery defect frames cannot show).
- If two pacing retakes fail, change the dialogue length/structure, not just the adverbs.
Multi-shot film assembly
- Each shot: one prepared first/last composition (the SAME full scene at both endpoints), one native 10s clip, automaticRetries=0, review gate between takes.
- Join with short dissolves only (e.g. 10-frame fade at each 10s boundary, outgoing endpoint held); native dialogue stays in its original interval. This is assembly, not repair compositing — confirm which the user's rule targets before refusing or proceeding.
- After every reassembly, verify master sources by hash, not by file presence: a rebuild can silently use a superseded clip for an "unchanged" slot (caught once by sha256 comparison + transcribing the assembled master's first interval).
- Transcribe the assembled master's per-shot intervals as the final acceptance gate; the whole-film silence map should show only deliberate boundary pauses and in-shot sentence breaths.
Rules
- Freedom belongs in story, pacing, atmosphere, metaphor and materials; exact brand artwork, factual UI and current product capabilities stay protected. State the balance explicitly when briefs iterate: identity/product geometry fixed, model free in ambient motion, glass/reflections and decorative particles.
- Keep dialogue content unique across sibling shots of one film: scan every shot's spoken line against the others before paying (a self-introduction duplicated by the closing line forced a paid retake). First/last-frame control preserves layout but does NOT preserve card text content through the middle of the clip — review mid-clip frames, not just endpoints.
- Test contrast for each logo, including all its component colors. A shared palette need not mean identical plaque backgrounds.
- Do not hardcode any one brief's palette, architecture, shot length or card layout as a universal style. These were one brief's choices.
- Respect the user's production method. If local compositing is excluded, do not silently add it as a fidelity workaround. Explain limits if native generation cannot meet exactness. Boundary joining of independently generated clips with short dissolve transitions is assembly, not repair compositing.
- Confidentiality: project media, real dashboards, prompts containing product names and learning/history data may be company confidential. Keep them in the project's private folders and experience docs; public skill copies carry generic examples only. Never expose API keys or include them in artifacts. Authorization to design does not authorize unlimited paid retries or later shots.
- For perspective cards or talking mascots, prepare the intended geometry and mouth in the input reference. HTML/CSS is useful for producing that image; pasted code is not executed by the video model. Keep speaking-character instructions consistent with the visible design.
- Native edits and speaker references can fail with generic content_blocked even for benign project media. Record exact inputs and errors, do not generalize the failure into a universal ban, and do not promise editing success.
Common mistakes
Independent logo references being redrawn; full dashboards acquiring invented UI; flat white panels undermining cinematic styling; low-contrast tone-on-tone; identity and style references being confused; generic automated QA passing visible errors; regenerating after a download timeout; short greeting lines leaving dead air; trusting endpoint frames when mid-clip card text drifted or speech stuttered; contradictory endpoint demands (blank signature region at BOTH frames while requiring the signature to persist); a revised brief silently dropping a required element the user had asked back in (always diff the new prompt against the rejected one); a reassembly silently pointing at a superseded source clip.
Validation
Check the prepared input first, then actual output. Label sampled inspection honestly; do not promise pixel identity. Record user-approved final status separately from model QA. Verify audio by silencedetect + transcription, and state plainly what was not listened to (speaker identity, pronunciation continuity).
References
- observed-api.md — observed Interactions API mechanics (recheck official docs before execution).
- example-welcome-film.md — generic, company-free walkthrough: 3×10s mascot+UI film, full prompts, pacing blocks, verification numbers, assembly config.
- Project-specific case studies live in each project's own private experience folders — never in this skill.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

