A Claude skill to generate video: what works today
Anthropic's pre-built skills cover documents, not video. A Sume CLI skill lets Claude Code make avatar videos; other clips need MCP or the API.

Anthropic's pre-built Agent Skills don't make video: they cover PowerPoint, Excel, Word, and PDF. A skill is instructions plus scripts, so a video skill needs a video tool behind it. For Claude Code, the current Sume CLI bundles skills you can install; one of them, sume-avatar-video, teaches Claude to create talking avatar videos through the CLI, behind its --confirm-paid gate for paid work. For text-to-video clips, connect Sume's hosted MCP server or call the API.
Anthropic's side comes from its Agent Skills overview and Claude Code's skills docs; Sume's from the CLI docs on Agent skills and Generation workflows, all read on 2026-09-27, plus the CLI's current source for the bundled skill files. Sume's basics page says the CLI still works but is not the primary integration path today.
What does Sume's avatar video skill do?
In the current CLI, its description says it creates, batch-plans, watches, and inspects Avatar Video 1.0 jobs from selected avatar handles, with paid gates. It starts from a ready avatar, which the sume-avatar skill creates or selects. The skill has Claude read the command's schema first, run sume avatar-videos create with a script, an avatar handle, and --confirm-paid, then follow the job with sume jobs watch and read the result.
Two of its rules matter when a render is slow: a watch that times out does not end the job, and Claude must not submit a duplicate paid job while recovering. Its own text also says not to route generic video generation through it.
How do I install it for Claude Code?
- Install the CLI and sign in with
sume login; How to generate avatar video from the terminal covers both. - Check the project folders. In current code, install writes to
.agents/skills/when the project has a.agents/folder, otherwise to.claude/skills/when it has.claude/, and stops with an error when it has neither. Claude Code's docs list.claude/skills/for project skills and~/.claude/skills/for personal ones, so run the install where.claude/exists and.agents/doesn't. - Review the files with
sume skills export <name> --json; without--json, current code prints only the skill name and file count. Then install the avatar and avatar video skills by name. - Start Claude Code in the project. Claude loads a skill when a request matches its description, or you run it as
/sume-avatar-video, since the folder name becomes the command.
mkdir -p .claude
sume login
sume skills export sume-avatar-video --json
sume skills install sume-avatar
sume skills install sume-avatar-videoWhat can the skill make, and what needs another route?
A skill can only drive what its tool does, and the CLI submits avatar work only. The CLI docs still point image and video work at Image 1.0 and Video 1.0, which are retiring; new integrations use POST /v1/images and POST /v1/videos with sume/auto, as Video 1.0 and Image 1.0 are retiring explains. In Claude Code, Sume's hosted MCP server gives Claude generate_video instead; see Can Claude make videos?.
| Family | CLI submit today | Other route |
|---|---|---|
| Avatar 1.0 | sume avatars create | The API, or hosted MCP avatars_create |
| Avatar Video 1.0 | sume avatar-videos create | The API, or hosted MCP avatar-videos_create |
| Video | No CLI subcommand | POST /v1/videos, or hosted MCP generate_video |
| Image | No CLI subcommand | POST /v1/images, or hosted MCP generate_image |
What are the limits?
- Avatar video scripts must estimate to 4–60 seconds, inclusive, and the CLI checks this before it submits.
- In current code, the avatar speaks English only.
--confirm-paidmarks generation that can reserve or spend credits, and Sume's docs say agents should not submit paid work without explicit operator confirmation. Rates are on API pricing.- Anthropic advises using skills only from trusted sources and auditing their files;
sume skills export <name> --jsonprints them for review.
Sources
Related posts
More in Agents
- Claude text to speech: make audio files with an MCP tool
For text to speech in Claude, connect a text-to-speech tool over MCP: Claude sends the script and a voice, then hands back a link to the audio file.
- Generate images for a website with Claude Code and Sume MCP
Connect Claude Code to Sume's hosted MCP with Write access, preview the cost with dry_run, then generate site images and save them in your repo.
- How to make a video ad with AI: brief, draft, approve, reuse
To make a video ad with AI, brief a video agent on the deliverable, review a draft before the paid render, then save the recipe to reuse it.
- Human in the loop AI agents: where a person approves
Human in the loop means an AI agent stops for a person before steps that matter. How Sume's agent does it in chat, and what replaces it in API runs.
Written by Sume