veo-render
Skill0 starsChanged 3 months agoMITView on GitHub
Generates video shots with Google Veo on Vertex AI from a shot list — handling authentication, cost confirmation, submission, long-running polling, retries and download, and resuming a partly-rendered batch. Use when rendering or re-rolling shots for a film, when a Veo job fails or is filtered, or when setting up Google Cloud access for video generation.
- Reads credentials
In this repository6 other files it gives its agents.
- final-cutSkill
- make-filmSkill
- score-filmSkill
- shot-designerSkill
- shot-reviewSkill
- story-architectSkill
The file
We split this skill into the rules it gives an agent
4 rules. About 0 are common, 1 is distinctive, and 3 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 57
Before any batch, show the arithmetic:
- Very specific: line 106
Keep --max-parallel at 3–4.
- Very specific: line 133
Never set generate_audio: true.
- Very specific: line 136
Keep every take, including rejected ones, in shots/rejected/.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Veo render
Turn the veo shots in shotlist.json into files in shots/. Everything here is
${CLAUDE_PLUGIN_ROOT}/scripts/veo_client.py; this skill is about using it well.
Auth
The client looks for credentials in this order and stops at the first that works:
$GOOGLE_ACCESS_TOKEN— a short-lived OAuth tokengcloud auth print-access-token- Application Default Credentials via
google.auth
In a cloud sandbox, use option 1. There is no gcloud here and a
service-account key file should not be written into an ephemeral container. Ask the
user to run this on their own machine and paste the result:
gcloud auth print-access-tokenThen, here:
export GOOGLE_ACCESS_TOKEN='ya29....'
export GOOGLE_CLOUD_PROJECT='their-project-id'The token lasts about an hour. Long batches will need a fresh one; the client fails loudly and resumably rather than half-writing a shot, so re-exporting and re-running is always safe.
Prerequisites on their project, once: the Vertex AI API enabled, and their account
holding roles/aiplatform.user. If a call returns 403, that is almost always one
of those two.
Model and cost
| model | ~$/s | when |
veo-3.1-fast-generate-001 | ~0.15 | default. All exploration, most shots |
veo-3.1-generate-001 | ~0.40 | hero shots only, once the look is proven |
veo-3.0-generate-001 | ~0.40 | fallback if 3.1 is unavailable in the region |
Rates are indicative and move — the client's table is marked to be verified in the
user's own console, and --rate or $VEO_RATE overrides it.
Always explore on Fast. Get the composition, the camera move and the light right at a quarter of the price, then re-render the two or three shots that carry the film at full quality with the same prompt and seed.
Before any batch, show the arithmetic:
python3 "${CLAUDE_PLUGIN_ROOT}/scripts/veo_client.py" estimate \
--shotlist shotlist.json --model veo-3.1-fast-generate-001It prints the one-pass cost and the realistic 3× regeneration cost. Give the user the second number. Budget 3–4× final runtime in generated seconds — about one shot in three needs another roll, and hero shots often take three.
Running
One shot, while you are iterating on a prompt:
python3 "${CLAUDE_PLUGIN_ROOT}/scripts/veo_client.py" generate \
--shotlist shotlist.json --shot 03 --outdir shots/ \
--model veo-3.1-fast-generate-001The whole film:
python3 "${CLAUDE_PLUGIN_ROOT}/scripts/veo_client.py" batch \
--shotlist shotlist.json --outdir shots/ \
--model veo-3.1-fast-generate-001 --max-parallel 3 --yesbatch is resumable: a shot with an existing .mp4 is skipped unless --force.
Re-rolling one rejected shot after review is --only 03 --force.
Both commands fold shotlist.look into every prompt automatically — the suffix,
the global negative and the style reference images. You never paste the look into
a prompt by hand, which is precisely why it stays identical across the film.
Generation takes minutes, not seconds — typically 1–3 minutes per 8s clip, and the client polls for up to 15. In an agent sandbox the shell tool will time out long before that, so launch detached and poll:
nohup python3 "${CLAUDE_PLUGIN_ROOT}/scripts/veo_client.py" batch \
--shotlist shotlist.json --outdir shots/ --yes > /tmp/veo.log 2>&1 &Then check /tmp/veo.log and shots/render-report.json periodically. Do not sit
in a tight polling loop; check every minute or two and do something useful in
between — render the motion graphics, cut the animatic.
Keep --max-parallel at 3–4. Quota is 50 requests/minute on GA models and 10 on
preview, but the real reason is that a bad look bible caught after three shots
costs less than after twelve.
When it fails
raiMediaFilteredReasons in the response. Safety filtering. Usually a person
described too specifically, something read as a real brand, or a scene that reads
as violent or medical. Rewrite the subject more abstractly — "a lone engineer"
rather than a described individual — and re-run.
403. API not enabled, or the account lacks roles/aiplatform.user.
404 on the model. That model is not available in that region. Try
--location us-central1, then a veo-3.0 model.
429. Lower --max-parallel. The client retries with backoff.
Token expired mid-batch. Re-export and re-run the same command; finished shots are skipped.
The clip is fine but wrong. That is not an error, that is shot-review.
Hygiene
- Record the
seedof every approved shot back intoshotlist.json. Without it a take you liked is not reproducible. - Never set
generate_audio: true. The film has a scripted voice. - Generate 8s and trim in the edit. The extra headroom is worth more than the saving.
- Keep every take, including rejected ones, in
shots/rejected/. Twice in a typical film the "worse" take turns out to cut better.
References
references/vertex-api.md— the raw REST surface: endpoints, request shape, every parameter, polling, GCS output, and curl equivalents for debugging
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

