Agent skills
Sume CLI ships agent guidance in two forms:
sume initinstalls the default project instructions.sume skillsinstalls focused Sume skill packs from the public registry.
Both are designed to route agents toward CLI/MCP schemas instead of copying large static API schemas into prompts.
Install default project instructions
Targets:
| Target | Output |
|---|---|
claude | .agents/skills/sume/SKILL.md when .agents/ exists, otherwise .claude/skills/sume/SKILL.md. |
cursor | .cursor/rules/sume.mdc. |
agents-md | Managed <!-- BEGIN sume:sume --> block in AGENTS.md. |
all | All supported targets. |
In a blank project, create an agent root first:
Skill registry
List available skills:
Install a skill:
Update installed skills:
Remove a skill:
The default registry is:
The CLI verifies registry file checksums before writing installed skill files.
Available Sume skill packs
| Skill | Use for |
|---|---|
sume | Top-level setup, auth, safety, update, and workflow routing. |
sume-assets | Asset Library search, uploads, downloads, and scene clip semantics. |
sume-brand | Brand DNA read/analyze workflows and reference image handling. |
sume-generation | Image/video generation, async jobs, polling, results, and downloads. |
sume-ads | Paid Ads video, avatars, Face Swap, and source-video boundaries. |
sume-reference-analysis | Public Reel/TikTok reference analysis and result summaries. |
sume-tools | MCP setup, schema discovery, skill refresh, and tool routing. |
What skills should tell agents
Sume skills should instruct agents to:
- start read-only with
sume --version,sume me --json, andsume credits --json; - use
sume login --no-browser --timeout 600in remote/headless terminals; - inspect
sume tools list --jsonandsume tools schema <tool> --json; - ask before running paid/provider generation;
- avoid printing secrets, signed URLs, full private media URLs, provider payloads, user ids, workspace ids, and raw emails;
- prefer CLI/MCP schemas as the source of truth for exact parameters.
Hermes, Modal, and remote terminals
When an agent is running in a remote terminal, it should not run plain
sume login and assume a local browser exists. Use:
For agent-controlled terminals, use the background login pattern from Authentication, show the URL/code only to the requesting user, then verify with:
Do not ask the user to paste API keys into chat unless the user explicitly chooses that credential path.
Registry overrides
Use overrides only for local registry testing:
Production users should use the defaults.