Google Omni
Full Editing Guide
Learn what Gemini Omni Flash can do, how to prompt it, and how to edit existing video without losing what already works.
Start here
What Is It? What's Special?
Google Gemini Omni Flash is a multimodal AI model that can turn text, images, audio, and video into short video clips with synchronized sound. What makes it special is that it understands and generates every part of the experience together, rather than treating visuals, motion, and audio as separate steps.
Accepts text, images, audio, and video in a single prompt
Generates video with native dialogue, music, ambience, and sound effects
Responds closely to technical camera instructions such as push-ins, orbits, and locked-off shots
Uses world knowledge to create more realistic physics, historical details, and scientific explanations
Supports multi-turn editing so you can refine one element while preserving the rest
Combines fast generation with strong visual and audio quality
Works best with a structured six-dimension prompt covering camera, style, lighting, location, action, and text
Can be used to edit pre-recorded video
Access and usage
How To Use It?
You can use it for free at labs.google. At the moment, you get 50 credits per day per free Google account. Depending on the specifications of the video, this is enough for several videos a day. If you have multiple Google accounts, you can get basic usage through this method.
If you want more usage, you can subscribe to Google's AI subscriptions.
You can also use this model through various API gateways, such as Kie.ai, OpenRouter, and Fal.
This allows you to have usage on a cost-per-generation basis. It also allows you to connect the model to your AI agents or software tools for programmatic usage.
Create better generations
Prompting Guide
The best way to prompt Gemini Omni Flash is to give it a concise production brief rather than a vague description. Cover the six key dimensions: camera, style, lighting, location, action, and text or audio.
Define the shot and camera movement
Explain the framing and movement: close-up, wide shot, locked-off camera, push-in, orbit, handheld, or one continuous shot.
Describe the visual style
Specify the aesthetic, such as cinematic 4K, documentary, film noir, watercolor animation, product commercial, or architectural visualization.
Set the lighting and location
Include the time of day, light quality, setting, weather, and environment. Specific locations usually produce stronger results than generic ones.
Describe the action in sequence
Use clear, verb-first instructions and include timing. Explain what happens first, what happens next, and how the scene ends.
Direct the sound and text
State whether you want dialogue, music, ambience, or sound effects. For on-screen text, provide the exact words, appearance, color, and animation.
Edit one thing at a time
After the first generation, make focused changes such as “Change the lighting to blue hour” or “Make the camera movement slower.”
Preserve successful elements
End every revision with: “Keep everything else identical.”
Edit existing footage
Editing Prompting Guide
For video-to-video editing, treat the uploaded clip as the source of truth. Clearly describe what should change, what must remain unchanged, and the desired visual or audio result.
Identify the source video
Refer to the uploaded clip as the source. Describe the important subject, setting, camera angle, and existing action.
State the exact edit
Ask for one focused change at a time, such as changing the weather to rain, replacing the background with a studio, changing the clothing color, adding subtle cinematic lighting, or removing an object from the scene.
- • Change the weather to rain
- • Replace the background with a studio
- • Change the clothing color
- • Add subtle cinematic lighting
- • Remove an object from the scene
Protect what already works
Explicitly preserve the subject’s identity, movement, timing, framing, camera motion, and composition.
Describe the target style
Specify the look you want: realistic, cinematic, documentary, vintage film, animation, product commercial, or another clear aesthetic.
Give audio instructions separately
Say whether to preserve the original audio, replace it, remove it, or add specific ambience, dialogue, music, or sound effects.
Avoid stacking too many changes
Make the first edit, review the result, then make another focused change in a new turn.
Use a preservation phrase
End the prompt with: “Keep everything else identical to the source video.”
More examples
You can see more examples in Google's official Gemini Omni prompt guide.
Installable workflow
The Omni Skill
The Omni Skill accepts a video and turns it into a strong, structured Omni prompt. It helps you describe what is already in the footage, decide what to change, and preserve the details that are working.
To install it in Claude, go to Customise, then Skills, choose Add new, and paste in the skill text below.
Show the skill text
---
name: omni-edit-prompt
description: >
Turn a raw video into a ready-to-paste editing prompt for Google Omni. Transcribes the
video with Deepgram (word-level timestamps), then writes a detailed edit prompt that
cites exact quoted lines and timings and calls out whip pans, camera angle changes,
custom on-screen typography, overlays, motion graphics, and sound effects. Use when the
user says "/omni-edit-prompt", asks to turn a video into an Omni edit prompt, or wants an
editing brief/prompt generated from raw footage.
argument-hint: " <path-to-video>"
allowed-tools: Bash, Read, Write, AskUserQuestion
user-invocable: true
---
# /omni-edit-prompt - Raw Video to Google Omni Edit Prompt
Pipeline: raw video → Deepgram word-level transcript → analysis of the script's beats →
a detailed, timestamped editing prompt for Google Omni.
---
## Step 1 - Get the video
If the user gave a path in their invocation arguments, use it. Otherwise ask for the path
to the raw video file. Confirm the file exists before continuing.
## Step 2 - Deepgram API key (first-run gate)
This skill needs its own Deepgram API key. Do not silently reuse a Deepgram key found
elsewhere, such as a global DEEPGRAM_API_KEY for other skills. Check, in order:
1. \${CLAUDE_SKILL_DIR}/.deepgram_key - a local file from a previous run of this skill.
2. If it doesn't exist, this is the user's first time running this skill. Ask them directly
in chat for their Deepgram API key. They can generate one free at
https://console.deepgram.com/. Explain it's needed to get word-level timestamps for
the transcript.
3. Once provided, save it with the Write tool to
\${CLAUDE_SKILL_DIR}/.deepgram_key (no trailing newline) so future runs don't ask again.
Do not print the key back to the user or log it anywhere else.
Never hardcode a key in this skill's files or commit .deepgram_key.
## Step 3 - Transcribe with word-level timestamps
Run:
\`\`\`bash
python3 "\${CLAUDE_SKILL_DIR}/scripts/transcribe.py" \\
--video "<path-to-video>" \\
--api-key "$(cat "\${CLAUDE_SKILL_DIR}/.deepgram_key")" \\
--out "<video-dir>/<video-name>.transcript.json"
\`\`\`
This calls Deepgram (model=nova-2, smart_format, punctuate, paragraphs, utterances)
directly on the raw video and writes a JSON summary with:
- transcript - full text
- duration - video length in seconds
- words - [{word, start, end}] for every word, second-accurate
- paragraphs - sentence-level chunks with start/end times
If the script reports the direct video upload failed, it automatically retries by
stripping audio with ffmpeg first. No action is needed unless both fail, in which case
report the error to the user.
Read the resulting JSON file to work from.
## Step 4 - Analyze the transcript for editing beats
Read through words and paragraphs and identify, with exact timestamps:
- The hook - the first ~3-5 seconds; what's said and how it should land visually.
- Topic/beat changes - sentences where the subject shifts, which are good spots for a
whip pan or hard cut.
- Lists or multi-part statements, such as "three things" or "first... second... third" -
good spots for camera angle changes per item, one per beat.
- Emphasis words or punchlines - numbers, surprising claims, and key terms - good spots
for custom typography call-outs or a zoom punch.
- Names, stats, quotes, or claims worth putting on screen - good spots for lower-third or
overlay graphics.
- Abstract or hard-to-visualize concepts - good spots for a motion graphic or animated
diagram.
- Pauses, cuts, or high-energy moments - good spots for a sound effect such as a whoosh,
pop, riser, or impact.
- The ending/CTA - how the last line should resolve visually.
Every beat you call out must be backed by the actual words array. Quote the exact phrase
(3-8 words) and give its precise start timestamp in mm:ss, converting from seconds. Do
not invent moments that aren't in the transcript.
## Step 5 - Write the Google Omni edit prompt
Produce a single prompt, written as direct instructions to Google Omni, structured like
this:
\`\`\`
EDIT BRIEF: <short title based on the video's topic>
Source video: <filename> | Duration: <mm:ss>
STYLE
<1-3 sentences on overall pacing/energy/tone for the edit>
TIMELINE
[00:00-00:04] "<exact quoted line>" - <specific edit: whip pan / camera angle change /
custom typography reading "..." / overlay / motion graphic / sound effect, and why>
[00:04-00:09] "<exact quoted line>" - <specific edit instruction>
... one entry per beat identified in Step 4, in chronological order ...
TYPOGRAPHY
<Any recurring on-screen text style guidance: font weight/feel, where key words should
pop up as kinetic text, which numbers/stats get their own card>
SOUND DESIGN
<A consolidated list of every sound effect cue with its timestamp: e.g. "00:07 - whoosh
on whip pan", "00:22 - riser into punchline at 00:24">
OUTPUT
<Aspect ratio / format guidance if known from context, otherwise omit>
\`\`\`
Requirements for this prompt:
- Every timeline entry must cite a real quoted phrase and timestamp from the transcript.
- Cover the whole video, not just the hook.
- Include at least one whip pan or camera angle change, one custom typography moment, one
overlay, one motion graphic, and sound effects timed to cuts or punchlines. Place them
where the transcript actually supports them, not evenly spaced for the sake of it.
- Keep instructions concrete and directive. Tell Omni what to do rather than simply
describing what the footage currently looks like.
## Step 6 - Deliver
Save the prompt to <video-dir>/<video-name>.omni-prompt.txt and show it to the user in
the chat response so they can copy it straight into Google Omni.
Build better AI content workflows
Learn the practical systems behind AI video, content creation, marketing, and automation.