Google Flow

Google Flow / Veo Playbook

Extended notes and the prompt that worked first try: flow-prompts.md in this repo.

Flow has no public API and no official MCP server, so generation runs through the browser. Veo itself is available through the Gemini API and Vertex AI if you ever want a scriptable path, but this kit does not use it.

UI map (assistant-driven projects)

• Gemini Omni Flash chat (right panel) = where generation happens. It replies with intent, then raises an Approve/Reject credit dialog.
• + → Create Scene = timeline assembly of existing clips only, no generation.
• Tools = community tools gallery, not the core generator.
• Videos list view shows every generation's full prompt; use it to diagnose old clips.
• Finished clips: open clip editor → download icon → lands in ~/Downloads.
• Uploads usually originated on this machine, so check ~/Desktop and ~/Downloads before re-downloading from Flow.

Credit hygiene (hard rules)

• Never approve a spend the user hasn't authorized. Veo 3.1 Lite ≈ 12 credits per 8s 720p/24fps clip (with audio).
• Draft at 360p first. Flow drafts at a lower resolution for a fraction of the credits, then upscales the version worth keeping (Aug 2026). Iterate on mood there and spend the full render once, on the clip you have already looked at. Note (Sep 2026): the assistant reports Veo 3.1 Lite as 720p-only at 10 credits; the 360p draft it offers is a different model (Omni 1.1 Flash, 6 credits). Ask for it by name if a draft is what you want, and say so in the approval you collected.
• Omni 1.1 Flash (Sep 2026): asked for ONE clip, the assistant proposed nine for 135 credits; reject and restate "exactly ONE video, 1 generation". It then quoted 12 credits and the dialog charged 15. The clip came back 10s, 720p, with a near-silent audio bed, so plan on MUSIC. The 1080p download is an upscale that runs a few minutes and lands in ~/Downloads on its own as a plain .mp4.
• Downloads arrive as a zip whose inner filename carries a non-ASCII character; unzip on macOS fails with "Illegal byte sequence". Extract with Python's zipfile.
• If the clip is brighter than the type can stand, tint it locally before spending again: a 65% navy color overlay in ffmpeg turned a cream highlight into a mid-navy that white Montserrat reads on, for zero credits.
• Export the final at 1080p, not 720p and not 4K: the recap composites at 1080x1920 or 1920x1080 depending on the project's ASPECT, so 1080p removes an upscale and 4K is thrown away.
• Read the assistant's confirmation before approving: model, count, duration, and aspect ratio. It has misread one ratio as the other, so ask for the one the project needs by name, "9:16 PORTRAIT (vertical)" or "16:9 LANDSCAPE (widescreen horizontal)", and correct it if the confirmation disagrees.
• Stale dialogs: consumed Approve/Reject dialogs linger in chat history and look clickable but are inert. Only a freshly-raised dialog does anything. If clicks change nothing, the dialog is dead; check whether the generation already ran.
• Prefer reusing existing clips (loop 8s → any length with a 1s crossfade) over generating.
• Start and end frame control (Aug 2026) can make an 8s background loop seamlessly, so the crossfade stops being load-bearing. Its headline use, holding a character consistent across a cut, is the one thing this kit does not do: any frame you supply must be abstract light only. A photo, a face or a rendered card as a start frame is rule 5.

Prompt rules (learned from failures)

1. No hex codes. #F2DF74 in a prompt became ribbon text "Best #2DF74".
2. No text you want rendered. Quoting text is a coin flip; render type locally instead.
3. Ban text redundantly when you want none: "absolutely NO text, NO letters, NO numbers, NO words, NO typography, NO logos, NO symbols".
4. Describe colors in words ("deep dark navy blue"), camera as "completely static" for backgrounds, and request the audio explicitly ("soft uplifting instrumental corporate background music, no vocals"); the audio bed comes free.
5. Nothing exact in-frame: faces, logos, and photos with text must never be AI-animated; composite them locally over textless generated motion.
6. A theme is text. When a caller asks for a background "matching the theme", steer only the colour, light direction and movement, and keep the theme wording out of the prompt entirely. In toastmasters-flow-promos that steering value is BG_MOOD in meeting.conf; substitute it for the palette clause of the background prompt and keep every NO. A prompt carrying "FRESH START" is how you get FRESH START rendered in wobbly letters across the backdrop, which is rule 2 with a friendlier face.

Verify every clip before using it

Contact-sheet it:

ffmpeg -i clip.mp4 -vf "select='not(mod(n\,24))',scale=640:-1,tile=4x2" -frames:v 1 sheet.png

Scan every second for surprise text or objects. One clip that looked clean grew the words "WHERE LEADERS ARE MADE" out of a paint stroke at the 6-second mark.