Google Flow Scripting

Google Flow Scripting and Prompting

Ask for surface, model, mode, and segment length before writing scenes. Google Flow and the Gemini API expose different capabilities. In Flow, Veo 3.1 commonly uses 4-, 6-, or 8-second clips, Veo Quality is more restricted, and Gemini Omni Flash supports 4-, 6-, 8-, or 10-second generations. Confirm the active interface because features change.

Storyboard every segment as stills and approve them before video generation. Panels first, approval second, video third.

Hard storyboard rule: "A storyboard-generation prompt must command image generation in its first sentence, declare the exact output count and layout, prohibit planning responses, and enumerate forbidden invented actions."

Version: 1.6.0

Use this Skill to turn a video brief into a production-ready, continuity-aware package for Google Flow or the Gemini API.

Non-negotiable output rules

1. Every referenced character or location uses a canonical @PascalCase handle with no spaces or punctuation, for example @Kwame, @CafeLadies, and @SidewalkCafe.
2. Use the canonical handle every time the referenced asset is named inside a generated script, not only in the attachment list.
3. Do not use em dashes in generated scripts. Use a colon, comma, full stop, parentheses, or spaced hyphen.
4. A storyboard copy-paste prompt begins exactly with:

GENERATE THE STORYBOARD IMAGE NOW.

5. A ten-second scene with four authored moments uses exactly four panels in a 2x2 grid. Use 3x3 only when nine distinct frames were explicitly authored.
6. Keep internal operator notes, the copy-paste generation prompt, and the approval checklist in separate sections.

Operating principles

1. Do not assume guaranteed implicit memory between independent generations. Carry continuity through self-contained prompt details, references, saved frames, supported first/last-frame controls, extensions, edits, and state tracking.
2. Treat platform limits as dated facts. Read reference/FLOW-FEATURES.md before making a capability claim. When the current interface differs, trust the active interface and record the discrepancy.
3. Plan assets before segment prompts. Establish cast, recurring objects, locations, canonical handles, reference priority, and set anchors first.
4. Build plates from approved images, not identity prose. Text descriptions can recreate a different person.
5. Storyboard before expensive video generation. Check identity, geography, composition, state, and handoffs in stills.
6. Make each generation prompt usable on its own. Avoid backward references such as "the same person" or "as before" unless an attached reference or frame is named explicitly.
7. Use project policies, not universal creative laws. Exact one-second beats, no on-screen text, and generated audio are configurable.
8. Treat Omni as a separate workflow. It supports conversational editing and broader multimodal context, but it does not support every Veo control.

Workflow

1. Establish the production profile

Resolve:

• surface: Google Flow or Gemini API
• model: Veo 3.1 variant, Gemini Omni Flash, or another named model
• mode: text to video, first frame, first and last frame, references to video, video edit, or extend
• segment length
• aspect ratio
• format and visual treatment
• runtime and delivery platform
• audio policy
• text policy
• rights-sensitive people, brands, locations, or claims

Do not write timed scenes until the first four items are known.

2. Read the relevant craft files

Use:

• core/flow-orchestrator.md
• core/flow-story-architect.md
• core/flow-asset-manager.md
• core/flow-storyboard-director.md
• the closest file under formats/
• core/flow-continuity-auditor.md

3. Build the production package

Produce, in order:

1. brief and assumptions
2. runtime arithmetic and segment map
3. cast, object, environment, and state inventory
4. canonical handle and reference plan
5. storyboard contact-sheet package
6. approved final panel and handoff plan
7. self-contained video prompts
8. mechanical validation results
9. human continuity audit and unresolved risks

4. Prompt text policy

Video, reference-sheet, and ordinary image prompts begin with one explicit text policy:

NO TEXT IN THE IMAGE: do not render words, labels, captions, or watermarks.

or:

INTENTIONAL TEXT IN THE IMAGE: state the exact words, surface, and placement; no other text.

Storyboard prompts begin with the image-generation command, then place the text policy near the top.

5. Audio policy

For generated-audio workflows, include an AUDIO: block with dialogue, effects, ambience, music, silence, and exclusions when useful.

AUDIO: Intentional silence.

For visual-only previs or sound added in post, state that policy and validate with --no-require-audio.

6. Generate a storyboard package

Use the agent instructions in core/flow-storyboard-director.md, the canonical template in prompts/storyboard-contact-sheet.md, or the executable generator:

python scripts/generate_storyboard_prompt.py examples/storyboard-spec.json \
  --output storyboard-package.md

7. Validate

Example for Flow and Omni reference generation:

python scripts/validate.py storyboard-package.md \
  --segment-length 10 \
  --surface flow \
  --model omni-flash \
  --mode references-to-video \
  --beat-mode off \
  --no-require-audio

Example for a Veo video prompt:

python scripts/validate.py video-script.md \
  --segment-length 8 \
  --surface flow \
  --model veo-3.1-fast \
  --mode text-to-video \
  --beat-mode exact

The validator is a mechanical linter, not a semantic continuity model. After it passes, run the continuity audit.

Installation and updates

Install with Python, Bash, or PowerShell:

python scripts/install.py --tool cursor
./scripts/install.sh --tool cursor
powershell -ExecutionPolicy Bypass -File scripts/install.ps1 --tool cursor

Check and apply future release updates:

python scripts/check_update.py
python scripts/update.py

Downloaded and installed copies never update silently. The updater verifies checksums and tests, detects local changes, and creates a backup before replacement. Git clones should use git pull instead. See reference/UPDATE-GUIDE.md.

Safety and rights

Before using a real person, protected brand asset, sensitive claim, archival material, or documentary reconstruction, identify consent, licence, disclosure, and factual-verification requirements. Never present a generated reconstruction as authentic evidence.

Supporting resources

• reference/PLAYBOOK.md: production methodology
• reference/FLOW-FEATURES.md: dated capability notes
• reference/TEMPLATES.md: prompt and deliverable templates
• reference/FAILURE-MODES.md: evidence-tagged failure table
• reference/UPDATE-GUIDE.md: safe updates and rollback
• reference/RELEASE-PROCESS.md: future release procedure
• release-history/: permanent notes for every version