Skills

A skill is a markdown playbook (SKILL.md) the operator installs through the dashboard. Each skill carries a name, a one-line description, and a body of instructions and templates. The Claude Agent SDK lists every installed skill in the system prompt as a <name>: <description> line, and when a description matches what the user is asking for, the agent reads the SKILL.md body and follows its instructions literally — including any output-format templates, forbidden phrases, or pre-submit lints the skill defines.

Skills override the agent's prose instincts. The templates exist precisely because the LLM default would otherwise drift.

How they work

1. The operator opens /skills in the dashboard and uploads a SKILL.md.
2. The dashboard writes the skill to the runtime DB.
3. The worker materializes installed skills to ~/.claude/skills/<name>/SKILL.md inside the container's claude-home volume.
4. The Claude Agent SDK auto-discovers each one and surfaces it in the system prompt as <name>: <description>.
5. On the next turn, when the user's request matches a skill description, the agent uses its Read tool to load the SKILL.md body and follows it.

Skills carry content (instructions + templates). They do not carry capabilities — see below.

Capabilities are gated globally

A skill cannot grant the agent the right to read a file, run a shell command, or write to disk. Those capabilities (Read, Edit, Write, Bash, etc.) are gated globally under /settings/agent-capabilities, not per skill.

This is a deliberate inversion of the agent-skills.io default. In Zeno:

• The operator decides at the profile level which capabilities the agent has at all.
• Each skill assumes only those capabilities; it cannot expand them.
• A skill that needs Bash and the operator has Bash disabled simply cannot run that step. The skill's templates should be written defensively.

Authoring a SKILL.md

A skill file is plain markdown with YAML frontmatter:

---
name: deploy-summary
description: When the user asks for a deploy summary, produce a one-line verdict + counts + URL.
---

# Deploy summary

When invoked:

1. Use the GitHub connector to read the latest workflow run for the repo.
2. Count successful jobs vs failed jobs.
3. Reply with this exact shape, no praise adjectives, no markdown headers:
   `<verdict> · <successful>/<total> · <run-url>`
   where `<verdict>` is one of `green`, `partial`, `red`.

The description field is what the SDK will match against. Keep it specific enough to fire only on the intended user requests.

Where skills run

Skills run inside the container, on the operator's machine, with whatever capabilities the operator has enabled at the profile level. They never run on a server, never call out to a hosted skill registry, and never carry credentials of their own — credentials belong to connectors.

What skills are not

• Skills are not connectors. A skill tells the agent how to use existing capabilities. A connector adds new capabilities.
• Skills are not crons. A skill is invoked when the operator's message description matches. A cron is invoked on a schedule.