Google Flow: 7 Essential Agent Prompt Rules That Work

Google Flow prompt guide: brief the Agent before you generate, the prompt structure Veo 3.1 rewards, and where credits actually disappear.

Google Flow AI filmmaking and cinematic video generation

Most Google Flow tutorials teach you to write better video prompts. That is the wrong place to start, because the expensive mistake happens earlier.

Google Flow runs on credits. Every generation costs, and a clip that comes back with the wrong character, the wrong camera move or garbled on-screen text costs the same as one that works. The Flow Agent exists to catch those problems before you spend anything, and it is the feature most people skip past on their way to the prompt box.

This covers how to brief the Agent, how to structure prompts once you get there, and the credit-burn traps worth knowing about.

In this guide:

Google Flow Agent storyboarding a scene before generation
Planning with the Agent is free. Generating is not. That asymmetry is the whole argument for briefing before prompting.

What the Google Flow Agent Actually Does

Google Flow is Google Labs’ AI filmmaking platform, announced at I/O on 20 May 2025 and available at flow.google. It runs on Veo 3.1 for video, Imagen for stills, and Nano Banana for image editing.

The Flow Agent arrived in 2026 and is built on Gemini with access to your project context. It is not a prompt rewriter. It brainstorms scene ideas, generates variations for comparison, and asks clarifying questions before generation starts.

That last behaviour is the whole point. The Agent interrogating your concept costs nothing. A generated clip costs credits. Any ambiguity the Agent resolves in conversation is ambiguity you do not pay to discover in a render.

You can also give the Agent standing instructions and brand rules, which persist across a project and prevent recurring errors. On-screen text coming out garbled is the classic example, and a brand rule specifying exactly what text should appear fixes it once rather than every clip.

Brief the Agent Before You Prompt

The Google Flow workflow that wastes the least money looks like this:

Open a project and talk to the Agent first. Describe the film, not the shot. What is it for, how long, what tone, who is in it.

Let it ask questions. Answer them properly. This is the cheapest part of the process and the only stage where changing your mind is free.

Review the production plan. The Agent will propose a structure. Modify it here rather than after you have generated half of it.

Set brand rules and Agent instructions. Style, palette, any on-screen text, anything that must stay consistent.

Then generate. Flow shows a cost approval step before each generation, which is worth actually reading rather than clicking through.

Community-published enterprise workflows formalise this into a system prompt pasted into the Agent chat before describing the project, turning Flow into a planning environment rather than a generation box. That is more process than most people need, but the underlying idea holds: plan in conversation, generate once.

The Prompt Structure That Works

Google’s own guidance identifies the Google Flow prompt elements that matter. A prompt with all of them outperforms a longer prompt without them.

Element What to specify
Subject and action Who or what, and precisely what they do
Composition Wide shot, close-up, medium two-shot
Camera motion Tracking, dolly, pan, tilt, aerial, push-in
Location A described place, not just a named one
Lighting Golden hour, neon, overcast, practical sources
Audio Veo 3.1 generates native audio; direct it explicitly
Style Cinematic realism, stop motion, clay animation

Two examples in that shape:

Sweeping drone shot of a lone hiker crossing a fog-covered mountain ridge at dawn, cinematic realism, shallow depth of field. Audio: wind, footsteps, ambient birdsong.

Close-up of a barista pouring latte art in a warm café, slow motion, golden hour lighting. Audio: coffee machine, soft jazz.

Note that both name the audio. Veo 3.1 generates sound natively, and leaving it unspecified means accepting whatever the model decides. Directing audio costs no extra credits and materially improves the result.

Ingredients and Character Consistency

The Google Flow Ingredients system, inherited from Whisk, is how you keep a character looking like the same person across shots.

An ingredient is a reusable visual element: a character, a prop, a stylistic reference. Create one from a text prompt via Imagen, or upload your own image. Scene Builder then preserves that appearance across every shot as camera angle, lighting and environment change.

This is worth building before you generate any video. Reference images cost image credits rather than video credits, which are considerably cheaper. Establishing your character and key environments as stills first, then moving to video, is the single largest saving available in a Flow project.

Newer features extend this further: frame-to-video animates a key frame you already like, and you can request new camera angles from an existing clip rather than regenerating the scene from scratch. There is also targeted editing that fixes one detail without re-rendering the whole clip.

Seven Rules for Google Flow Prompts

1. Specific beats long. A rambling prompt with ten competing directions produces worse output than a short precise one. Flow rewards specificity, not complexity.

2. Always direct the audio. Native audio generation is on by default. Unspecified means unpredictable.

3. Change one variable at a time. Generate, alter only the camera move or only the lighting, generate again. Comparing two clips that differ by one element teaches you how the model responds faster than any guide.

4. Build ingredients before video. Stills are cheap, video is not. Lock the character and environment first.

5. Put on-screen text in brand rules, not prompts. Text rendering is the most common failure, and a standing rule fixes it across the project.

6. Use the Agent for coverage, not just ideas. Ask it what shots you are missing. Filmic coverage is a structural problem it handles well and humans routinely forget.

7. Read the cost approval screen. It exists so you can catch a mistake before paying for it. Clicking through defeats the one guardrail Flow gives you.

Where Credits Actually Go

The honest accounting, because this is what determines whether Google Flow is affordable for a given project.

Video generation is the expensive operation. Image generation for ingredients is much cheaper. Regenerating a full clip to fix one small element is the most wasteful thing you can do, which is why targeted editing and frame-to-video exist.

Reported end-to-end time for a short film runs roughly 30 to 90 minutes depending on how many iterations you need. Iterations are the variable that matters, and iterations are exactly what Agent planning reduces.

The pattern that wastes the most credits is generating first and deciding what you wanted afterwards. Every clip you throw away was paid for at full price.

Access and Regional Limits

Two practical constraints worth knowing before you plan around Google Flow.

Account type. Sign in with a personal Gmail account. Workspace accounts from a company or school have restricted access to some features, which is an odd limitation for a tool aimed partly at production teams.

Region. Availability covers 149 or more countries, with reported restrictions affecting some EU and UK users. Check before building a workflow around it.

Google Flow also sits inside Google’s wider model lineup rather than standing alone. Our guide to AI video generators covers how it ranks against Runway, Kling and the rest, and AI models compared covers the Gemini models underneath it.

Common Google Flow Mistakes

Four patterns that cost credits, drawn from how projects typically go wrong.

Generating before planning. The most expensive mistake and the easiest to avoid. Every clip generated before you knew what you wanted was paid for at full price and thrown away.

Building character consistency in prompts rather than Ingredients. Describing the same character in words across ten prompts produces ten slightly different people. The Ingredients system exists precisely because text descriptions do not hold appearance stable.

Regenerating whole clips to fix small details. Google Flow has targeted editing and frame-to-video for exactly this. Re-rendering a full clip because one element is wrong wastes most of what you paid.

Ignoring the cost approval screen. It is the only guardrail between a misread prompt and a spent credit. Clicking through it reflexively removes the protection.

How Google Flow Fits a Real Workflow

Worth being clear about where Google Flow sits, because it does not replace an editing suite.

Flow generates clips. It is strong at producing individual shots with consistent characters and directed camera work, and Scene Builder handles multi-shot sequences within the tool. What it does not do is finish a film: colour grading, audio mixing, titling and delivery still happen elsewhere.

The realistic pipeline is Google Flow for generation, then a conventional editor for assembly and finishing. Our coverage of DaVinci Resolve 21.1 is relevant here, since Resolve now ships an MCP server that puts that finishing stage under AI control too. Generation and post-production are converging on natural-language interfaces from opposite directions.

For anyone weighing Google Flow against alternatives before committing to a workflow, the ranked comparison in our AI video generators guide covers where each tool leads.

Frequently Asked Questions About Google Flow

Is Google Flow free?
There is free access, and generation runs on a credits system with paid tiers for more. The cost approval step before each generation is what keeps spending visible.

What model does Google Flow use?
Veo 3.1 is the default video model as of 2026, with Imagen for still generation and Nano Banana for image editing. The Agent itself runs on Gemini.

What is the Flow Agent for?
Planning before generation. It brainstorms, asks clarifying questions, proposes production structure and holds brand rules across a project. Its value is reducing the number of clips you pay for and discard.

How do I keep a character consistent across shots?
Use the Ingredients system. Create a character reference from a prompt or an uploaded image, and Scene Builder preserves that appearance across shots as the camera and lighting change.

Why does my on-screen text come out wrong?
Text rendering is a known weak point across video models. Specify the exact text in Agent instructions or brand rules rather than in individual scene prompts, which makes it consistent across the project.

Can I use a work Google account?
You can sign in, but Workspace accounts have restricted access to some features. A personal Gmail account is the recommended route.

Getting Your First Project Right

A concrete sequence for anyone starting from nothing, sized at roughly an hour.

Ten minutes, no generation. Open Google Flow, describe the project to the Agent, answer its questions, accept or amend the production plan it proposes. Set brand rules for style and any on-screen text.

Fifteen to twenty minutes on stills. Build your character and key environment references as images. These are cheap relative to video and they determine whether every later shot looks like the same film.

Generate one test clip. Not the opening shot, a middle one. See how the model handles your character and your camera direction before committing to a sequence.

Change one variable and regenerate. Adjust only the lighting or only the camera move. The difference between those two outputs tells you more about how Google Flow interprets direction than any amount of reading.

Then build the sequence. With a plan, locked ingredients and a calibrated sense of how the model responds, the remaining generations are far likelier to be usable first time.

The whole point of that order is that the free and cheap steps come first, and the expensive step happens once you know what you are asking for.

Verdict

Google Flow is a capable filmmaking tool, and the Agent is the part that changes how you should use it. Treating Flow as a text-to-video box and typing prompts until something works is how people burn through credits and conclude the tool is expensive. Treating it as a planning environment where you brief a co-director, agree a structure and then generate is a different experience at a fraction of the cost.

The honest caveat is that this is still generative video, and the limits are the category’s limits rather than Google’s. On-screen text remains unreliable, character consistency needs deliberate setup through Ingredients rather than arriving for free, and iteration count is the thing that actually determines your bill. None of the planning eliminates those; it reduces how many times you pay to hit them.

Concrete next step: before your next project, open Flow and spend ten minutes with the Agent generating nothing at all. Describe the film, answer its questions, get a production plan, set your brand rules. Then compare the credits that project consumes against your last one. That comparison is the whole argument for working this way.

Sources and Further Reading

For related coverage on this site, see AI video generators for how Flow ranks against the field, AI models compared for the Gemini models powering it, DaVinci Resolve 21.1 for the post-production side now under AI control, and AI content creation for the written end of the same workflow.

One comment

Comments are closed.