---

name: gemini-veo-studio description: Make short videos with Google Veo (Gemini API) from Claude. Opens a starter panel of video templates like the Gemini app, writes the prompt, generates clips with sound, then optionally finishes them in code (mascot swap, phone screens, captions, end card).

Gemini Veo Studio Skill for Claude

Claude plans and finishes; Veo films. Veo runs through the official Gemini API with the user's own key, so this works anywhere Claude has a shell (Claude Code, or a linked computer).

0. Setup (once)

• pip install google-genai
• The user needs an API key from https://aistudio.google.com/apikey, exported as GEMINI_API_KEY. Never print, log or commit the key; ask the user to set it in their own shell.
• Veo is not on the Gemini API free tier: it is billed per second of video on the user's Google account, and a Gemini app subscription does not cover it. Say this once before the first generation, with the prices in README.md. Check current prices at https://ai.google.dev/gemini-api/docs/pricing.
• If the network blocks generativelanguage.googleapis.com, run the script on the user's machine instead.

1. Show the starter panel first

When the user asks for a video and has no clear brief, show the panel before writing prompts:

• Run python3 scripts/panel.py and open or send panel.html. It is a gallery of templates (Creators, Brand, Restyle, Celebration), like the Videos page in the Gemini app. --thumbs makes real thumbnails with Gemini's image model, which is billed.
• In chat, offer the same choice with the question tool: first the category, then up to 4 templates from it. A free-text answer always works.
• Templates live in assets/templates.json. Each has {placeholders} (subject, place, line, name), the input it needs, and a default aspect. Add new templates there.

2. Write the shot list

• One Veo call = one shot of 4, 6 or 8 s. Plan 3 to 5 shots for a 15 to 40 s video.
• Prompt rules that work:
  • say the camera (locked-off, push-in, handheld);
  • say the light and the look;
  • put dialogue in double quotes;
  • describe sounds;
  • end with "No on-screen text" (Veo text comes out garbled; add text in edit).
• Same character in every shot: describe it identically, or pass up to 3 --ref images (8 s only).
• A phone or screen that will show a real UI: locked-off camera and a "flat solid chroma-key green" screen, so it can be replaced in edit.
• Save the shot list as a .md and show it before spending money on generation.

3. Generate

python3 scripts/veo.py "<prompt>" -o shot1.mp4 [--aspect 9:16] [--seconds 8] [--resolution 1080p]
python3 scripts/veo.py --template talking-pet --set subject="a cream cat" --set line="Hi!" -o pet.mp4
python3 scripts/veo.py "<prompt>" --ref mascot.png -o intro.mp4         # keep a character or product
python3 scripts/veo.py "<prompt>" --first a.png --last b.png -o x.mp4    # first and last frame
python3 scripts/veo.py "<prompt>" --extend shot1.json -o shot1b.mp4      # continue a clip (+7 s, 720p)
python3 scripts/veo.py --list | --dry-run

• Each run takes about 1 to 6 minutes.
• It saves the .mp4 and a .json with the request and the video URI. Extension needs that URI, and Google keeps generated videos for 2 days.
• In the EU, UK, CH and MENA, people generation is limited to adults.
• A clip blocked by safety filters returns no video and is not billed. Rephrase the prompt and retry once.
• Generate shots one at a time and look at each one: make a contact sheet with ffmpeg and read it before moving on.

4. Finish in code (optional, this is where Claude adds the most)

See references/post-production.md:

• join the shots;
• keep Veo's own audio;
• swap a stand-in character for the real mascot, with the mouth driven by the audio;
• replace green screens with a rendered UI flow;
• clean garbled text;
• add captions and an end card;
• export H.264.

5. Deliver

• Send the final .mp4 and say what is generated and what was added in edit.
• Mark example or illustrative numbers in the video itself.
• Do not claim a real person or brand endorsed it.