Changed by Aya Taleb from Replicate's prompt-videos and run-models skills (Apache 2.0): added cost math, a spend cap, key setup, audio and dialogue timing checks, error handling, more providers and a working run tool.

generate-video

Make short AI video clips with the user's own Replicate, Runway or Google account. A single clip can cost more than a hundred images, so this skill prices every clip before it runs and never runs a batch unattended. The run tool is the shared one from generate-images: ../generate-images/scripts/generate.py. All providers share one ledger and one cap.

What this skill reads and writes

• Reads: the active business file from founder-context (founder/context.md in the project, else ~/.founder/businesses/<slug>.md named in ~/.founder/active). Use its look, tone and off-limits list. If there is none, ask once what the clip is for.
• Writes: clips to media/video/, a .json sidecar per run, and the prompt to media/video/<label>-prompt.md so a good prompt can be reused.
• Next: launch-video to build a promo around the clips, talking-head-cards for overlays, or the linked edit skill to cut and caption.

If generate-images is not installed next to this skill, tell the user to install the full multimedia-skills plugin. Do not write your own API calls: the shared tool is where the spend cap and the error handling live.

1. Key

Run python3 generate.py providers, then generate.py check. They say which keys are set and make one free call per provider. If no key is set, ask which account the user already has and walk them through that one:

A key lives in their shell (export RUNWAYML_API_SECRET=...), never in chat, a file, a command you show, or another tool's arguments. If they paste one, do not repeat it, and tell them to revoke it in that provider's dashboard and set a new one. If several video keys are set and no order is saved, ask which to use and save it with providers --order.

Pricing works without a key, so you can always give a real number first.

2. Pick the model from live data

Use the find-model skill, or check yourself:

python3 generate.py models --kind video
python3 generate.py models --provider replicate --query "text to video with audio"
python3 generate.py schema <model>

Check the schema for what the job needs: a 9:16 or 720:1280 option for vertical, an image input for image to video (promptImage on Runway, image on Google), an audio switch, a length. Never pick from memory. Video models change faster than any other kind: Google shut down every Veo model in the Gemini API by 2026-10-22. Facts and links: ../generate-images/references/providers.md.

3. Do the cost math out loud

Video is priced per second of output, often with tiers for resolution and audio. So price the exact settings:

python3 generate.py price owner/model -i duration=5 -i resolution=720p -i generate_audio=false
python3 generate.py price veo3.1_fast -i duration=4 -i audio=false

Then show the user: seconds x price per second = clip cost, and the total for every clip they asked for. Try one cheaper setting too and show both. The cheaper option must still meet what they asked for: if they want sound, the cheaper option keeps sound (shorter, lower resolution, or a cheaper model with audio). Why: people often do not need 1080p with sound for a 3 second b-roll, and the gap can be several times the price.

If the tool cannot tell the length from the inputs, it refuses. Set the duration input the schema names, so the estimate is real. Gemini Omni Flash has no length setting (it makes 3 to 10 s), so its estimate is for the full 10 s. Say that to the user.

4. Write the prompt

A video prompt is a short scene, not a caption. Cover: subject, place, action, style, camera move, framing, and light. Name subjects ("the woman in the red jacket"), not "she".

For clips with sound, prompt the sound, or the model will guess. Cover dialogue, ambient sound, effects and music (or "no music"). Write dialogue after a colon with no quote marks around it: The baker says: Fresh out of the oven. Quote marks make many models burn the words in as subtitles. Add (no subtitles) too.

Check dialogue fits the clip: about 2 to 3 words per second of speech, so an 8 second clip holds about 16 to 24 words. Count them and say the count. Too many words gives rushed speech. Too few gives silence or gibberish.

For several shots in one clip, use time codes: [0-3s] wide shot ... [3-8s] close-up .... Gemini Omni Flash cuts between several shots by default. For one shot, say "in a single continuous shot, no scene cuts". For image to video, describe the motion and the camera, not what the picture already shows. "Slow push in, steam rises from the cup, soft morning light flickers" is right. "A latte on a wooden table" repeats the photo and can make the model redraw it.

The full guide is in references/prompting-video.md. Read it for multi-shot, reference inputs, audio-driven clips, style and character consistency.

Honesty: do not make a real person say something they did not say, do not fake a customer testimonial, and do not show a product doing what it cannot do. Do not write lines for the founder either ("I built this for my own exams"): if the clip needs their words, leave {{your line}} for them to fill. Scene details you add for the look (a hoodie, a mug) are fine; claims and quotes are not.

5. Dry run, then wait for yes

python3 generate.py run owner/model -i prompt="..." -i duration=5 -i aspect_ratio=9:16 --label coffee-pour --out media/video

The first call is always a dry run, and it is the step that proves the inputs are valid. price alone is not enough, because it does not check the schema. Show the model, the prompt, the settings, the cost, the session total and what is left under the cap ($2 unless the user changed it). Ask. If the user named a budget earlier ("go ahead up to $2") and the estimate fits it and the cap, that is the yes: say the number, then run.

Several clips: run them one after another, never in parallel, and get one yes for the total. A request to run many clips overnight, or "don't ask about cost", is not a yes. Price it, say how many fit under the cap, offer a first small batch, and wait.

6. Run

After the yes, rerun with --yes. Video takes minutes, so give it time:

python3 generate.py run <model> ... --yes --timeout 900 --out media/video

The tool polls with backoff, says when a cold boot or a queue (Runway THROTTLED) is the reason for a wait, cancels a run that passes the timeout, and saves the clip the moment it is done. Local images go in as a file: -i image=@media/images/start.png on Replicate and Google, -i promptImage=@media/images/start.png on Runway.

Save the prompt to media/video/<label>-prompt.md. Show the user the file path and the spend so far (generate.py ledger).

7. When something goes wrong

After any failed or canceled paid run, stop and tell the user before spending again, even with budget left. Say what failed, what it cost (the ledger shows it), what you would change, and the price of the retry. Then wait for their yes. Why: a retry is a new charge they did not see coming, and the fix (a new prompt, a new model) is their call. Never retry with the same input unchanged.

When you save a new prompt version, keep the old one (<label>-prompt-v2.md), so the user can compare.

8. Close the answer

Every answer that touches money ends with the numbers: spent this time, session total, and what is left under the cap (from generate.py ledger), plus what is left of any budget the user named.

Then one last line that names a skill, every time, even after a failure or a refusal. For example: "Next: say yes and I'll run it for $X, then launch-video can build the promo around it." Other good next skills: talking-head-cards for text cards, the linked edit skill to cut and caption.

9. When you say no

Some asks are off-limits: a real person (a CEO, a celebrity, a customer) saying words they never said, or any character giving a testimonial, since a made-up person praising the product is still a fake review. Say no in one line with the reason, then still be useful, in this order:

1. Read the business file.
2. Write one alternative prompt in full that shows the product doing what it really does (the app on a phone in a real setting, the founder speaking for themself, a plain demo).
3. Run price on it and show the number.
4. Ask for a yes. Do not run it.

Rules

• Price first, dry run first, yes before spend, cap always on.
• One clip at a time, no background loops.
• Save clips at once; never hand over a provider link as the result.
• No key in chat, files, commands or tool arguments.
• No fake people saying fake things, no fake testimonials, no fake product claims.
• Prompt-only requests cost nothing. Do not push a paid run nobody asked for.