Muse Google Veo Web

End-to-end video production with Google's Veo through the web interface (Google Flow) on a Google AI Pro subscription. This skill covers the full loop most web-based Veo projects need: setting up the account and project, writing prompts Veo responds to, driving generation in the browser, verifying every clip across its full duration, and stitching the result. It is backend-specific on purpose: the rest of the pack stays model-neutral, and this is where the Flow particulars live.

The workflow

1. Set up the project (references/flow-generation.md). Pro subscription, one Flow project per video project, one Google account for the whole job. Know your monthly credit allowance before you plan the clip count.
2. Plan the clip list. One clip per story beat, 8 seconds each: that is Veo's web sweet spot. Never stretch a handful of clips across minutes by looping; if the clip count gets heavy, trim the story, not the coverage.
3. Lock references first. Anything that must stay consistent across clips (characters, products, logos, locations) gets a reference sheet image before any clip is generated. Flow's Ingredients-to-Video mode takes reference images; attach them on every generation.
4. Write the prompts (references/veo-prompting.md). Five-part Veo formula: Cinematography, Subject, Action, Context, Style & Ambiance, plus Preserve/Avoid slots. For text-to-video, paste locked descriptions verbatim, never paraphrased. For Ingredients-to-Video, attach the reference and describe only the shot and motion; do not restate what the image already shows (repetition invites drift).
5. Generate in the browser (references/browser-ops.md). Queue clips sequentially in one session. Renders take a long time; a full batch runs for hours in the background, which is normal.
6. QA every clip across its full duration (references/web-qa.md). Sample frames across the whole clip, classify failures, fix or regenerate with an attempt budget, then cut what won't pass.
7. Download and assemble. Take the highest quality Flow offers, strip native audio if post handles the sound, and stitch with the assembly-and-delivery skill (scripts/concat.py).

Rules that don't bend

• Reference images on every Ingredients-to-Video generation. No ingredients, no consistency. The model does not remember across clips; you reattach every time. Pure text-to-video one-offs with no continuity burden are exempt.
• Full-duration verification. An 8-second clip is 8 seconds of chances to break. Check frames across all of it before it touches the timeline.
• One clip per beat. Looping a few clips across minutes reads as cheap the moment anyone notices, and they notice.
• Cheap tier for iteration, expensive tier for finals. Draft and explore on the cheap tier; spend the expensive tier only on clips that already passed QA. Verify current credit costs before budgeting; they change.
• Decide audio upfront. Either keep Veo's native audio (dialogue, ambience) or strip it and do all sound in post. Mixing both without a plan gives doubled ambience and sync drift.