How to Use Gemini Omni in VS Code: Prompt Files, Asset Workflows, and API-Ready Setup

By Sagnik Bhattacharya 20 May 2026 7 min read

Coding Liquids blog cover featuring Sagnik Bhattacharya for using Gemini Omni in VS Code, with prompt files, JSON manifests, asset folders, and video evaluation notes.

Using Gemini Omni in VS Code does not mean there is a magic Gemini Omni button inside the editor. As of this fact check, the reliable workflow is to use VS Code as your production workspace: prompt files, asset manifests, scene breakdowns, benchmark notes, and hand-off records, while generation happens in Gemini, Google Flow, or supported YouTube surfaces.

I teach Flutter and Excel with AI — explore my courses if you want structured learning.

That may sound less exciting than a fake one-click integration, but it is more useful for real work. It gives you version control, reusable prompt libraries, and a clean path to automation if Google exposes a suitable API surface later. If you are looking for Gemini CLI inside VS Code, read the Gemini CLI VS Code guide separately.

Note: This guide intentionally does not invent a Gemini Omni VS Code extension, package, endpoint, or API. It creates an API-ready folder structure and manifest system while keeping generation in official Google surfaces.
Note: Use the Flow guide for current access, subscription, and credit details.

Quick answer

Use VS Code to manage Gemini Omni prompts, references, scene manifests, review rubrics, exported file names, and version history. Generate in Gemini or Flow today; keep your project structured so it can plug into an official API later if one becomes available.

You create repeatable AI video prompts for clients, content calendars, or product videos.
You need prompt versioning and review records.
You want a practical workflow that can become API-ready without relying on invented endpoints.

Why VS Code Helps With AI Video

AI video looks visual, but the repeatable part is mostly text and asset management. Prompts, references, review notes, file names, licence notes, and edit history are all easier to manage in a code editor than in scattered chat boxes.

VS Code also gives you search, Markdown preview, Git diffs, snippets, JSON validation, and project folders. That makes it a good home for the work around Gemini Omni even when the actual generation happens in Google Flow.

Recommended Folder Structure

Start with a folder that separates prompts, assets, outputs, and evaluations. Keep source assets out of random downloads folders, because broken references are one of the fastest ways to lose repeatability.

gemini-omni-project/
  README.md
  prompts/
    001-product-hook.md
    002-dashboard-explainer.md
  assets/
    references/
    storyboards/
    audio/
  manifests/
    scenes.json
    assets.json
  evaluations/
    rubric.md
    results.csv
  exports/
    draft/
    final/

Prompt File Template

A prompt file should store more than the prompt. It should record the target model, aspect ratio, duration, references, and review criteria. That makes the same prompt rerunnable by you or another teammate.

The template below is designed for manual Flow use today and automated use later.

---
scene_id: 001-product-hook
model_target: Gemini Omni Flash
surface: Google Flow
aspect_ratio: 9:16
duration_seconds: 10
references:
  - assets/references/bottle-front.jpg
  - assets/audio/soft-click.wav
status: draft
---

Create a 10-second vertical product video of the reference bottle on a wooden desk. One continuous shot. Slow push in from medium shot to close-up. Morning window light from the left. The bottle rotates gently once. Preserve the bottle shape and label from the reference. No on-screen text.

Review targets:
- Bottle shape stays consistent.
- Label remains readable in the final three seconds.
- Motion is slow enough for a Shorts opener.
- No extra objects appear on the desk.

Scene Manifest For API-Ready Work

A manifest is not an API call. It is structured documentation that can become an API payload later. The benefit today is consistency: every scene has an id, a prompt, references, duration, aspect ratio, and review rubric.

If Google later exposes a documented Gemini Omni API path, this manifest can be transformed by a script. Until then, it keeps manual Flow work organised.

{
  "project": "q2-dashboard-video",
  "created": "2026-05-20",
  "surface": "Google Flow",
  "scenes": [
    {
      "id": "scene-001",
      "prompt_file": "prompts/001-product-hook.md",
      "duration_seconds": 10,
      "aspect_ratio": "9:16",
      "references": ["assets/references/bottle-front.jpg"],
      "evaluation": {
        "prompt_adherence": null,
        "motion_stability": null,
        "text_rendering": null,
        "audio_usefulness": null,
        "object_consistency": null,
        "editability": null,
        "cost_per_usable_clip": null
      }
    }
  ]
}

VS Code Snippets For Prompt Blocks

If you write many prompts, create snippets for camera, motion, reference, and review sections. This is faster than copy-pasting from old chats and safer than relying on memory.

Your snippets should encourage complete prompts, not decorative keyword stuffing.

Camera block: framing, move, speed, and whether the shot is continuous.
Reference block: which asset to preserve and which parts may change.
Constraint block: no extra objects, no text, simple background, preserve timing.
Review block: score outputs before deciding whether to edit, regenerate, or discard.

Evaluation Rubric In The Repo

Benchmarks are much easier when the rubric lives beside the prompts. For each generated clip, record the exact Flow cost, the active model, the output file, and the final decision. That gives you a real cost per usable clip instead of a vague impression.

Use the comparison framework in Gemini Omni vs Veo vs Seedance vs Sora if you are comparing multiple video tools.

Metric	Score question	Why it matters
Prompt adherence	Did the output do what the prompt asked?	Prevents beautiful but irrelevant clips.
Motion stability	Do objects warp, jump, or melt?	Catches expensive edit failures early.
Text rendering	Is text readable if requested?	Decides whether captions belong in the editor.
Object consistency	Does the subject stay recognisable?	Crucial for products and dashboards.
Editability	Can the clip be cut into a real sequence?	A clip needs usable handles, not just a nice middle frame.

Step-by-step: set up the VS Code workflow

This setup gives you a real working system even without a dedicated Gemini Omni VS Code extension. Generation still happens in Gemini, Flow, or supported YouTube surfaces; VS Code keeps the production work organised and reviewable.

Create the project folder. Use one folder per campaign or video series, not one folder for every random test.
Add the folder structure. Create prompts/, assets/, outputs/, evals/, and exports/. Keep raw references separate from generated clips.
Create a prompt file for each scene. Use names such as scene-01-hook.md and scene-02-dashboard-reveal.md. Each file should include goal, references, prompt, edit history, and review notes.
Create a JSON manifest. Store asset names, intended model surface, aspect ratio, duration, safety notes, and expected output filename. Treat it as a planning record, not as a fake API contract.
Preview the prompt in Markdown. Check that the subject, camera, lighting, constraints, and final-frame instruction are visible without scrolling through old chat history.
Generate in the official surface. Copy the prompt to Gemini or Flow, upload the listed references, and record the active model and credit cost in the manifest after generation.
Save outputs with version names. Use scene-01-hook-v01.mp4, scene-01-hook-v02-camera-fix.mp4, and similar names. Random download filenames kill repeatability.
Commit the prompt and review notes. Commit text assets and manifests. Do not commit large video files unless your repo is set up for that; store exported videos where the team normally keeps media.

Example review note

Add a short review block after every generation. This is what turns the VS Code workflow into something more useful than a prompt dump.

## Review - 2026-05-20
Surface: Google Flow
Model shown in settings: [record active model]
Aspect ratio: 9:16
Duration: 8 seconds
Prompt adherence: 4/5
Motion stability: 3/5
Reference preservation: 5/5
Text rendering: not requested
Decision: edit once
Next edit: Keep the product and lighting the same, but lock the camera and slow the rotation by half.

Optional VS Code snippet setup

If you create many Gemini Omni prompts, add a user snippet so every prompt starts with the same useful structure. In VS Code, open user snippets for Markdown and add a snippet like this:

{
  "Gemini Omni scene prompt": {
    "prefix": "omni-scene",
    "body": [
      "## Goal",
      "$1",
      "",
      "## References",
      "- $2",
      "",
      "## Prompt",
      "Create a $3 video. Subject: $4. Action: $5. Camera: $6. Style and lighting: $7. Constraints: $8. Final frame: $9.",
      "",
      "## Review",
      "- Prompt adherence:",
      "- Motion stability:",
      "- Reference preservation:",
      "- Edit decision:"
    ]
  }
}

This turns the editor into a repeatable prompt workspace. You can then link the prompt file to the Flow output and review note instead of rebuilding context from memory.

Common mistakes

Pretending VS Code is the generation surface when it is only the workflow surface.
Saving prompts without the model, duration, aspect ratio, or reference list.
Letting exported files keep random download names.
Treating a manifest as a real API contract before official documentation exists.
Skipping Git commits for prompt changes when a client or team needs traceability.

Sources

These official references are useful if you need the product or framework documentation alongside this guide.

Want to use AI tools more effectively?

My courses cover practical AI workflows, from spreadsheet automation to app development, with real projects and honest tool comparisons.

Browse AI courses