Skill

A Skill is a markdown-defined capability that gets injected into an Agent's prompt at run time. Skills are how you teach an Agent to use a specific tool, follow a specific protocol, or operate within a specific convention without re-writing the system prompt on every Agent.

What's in a Skill

A Skill row carries:

• slug and name — the slug is the identifier the daemon uses when listing skills in the prompt.
• description — a one-liner shown in the Skills picker.
• content — the markdown body. This is what actually gets appended to the Agent's prompt at run time.
• kind — either builtin (workspace-less, seeded by migration, auto-linked to every new Agent) or custom (workspace-scoped, opt-in per Agent).

Built-in skills

Two skills ship in the box and are auto-linked to every new Agent:

• memory — documents how the agent reads .evoxiv/MEMORY.md and the JSON-patch format for PATCH /agent/memory. Included in the primary author pass and the continuation pass that amends Memory after every run.
• story — documents the full Story-management surface for agents: POST /agent/stories, PATCH /agent/stories/:id, DELETE /agent/stories/:id. Used by investigative tasks that need to spawn follow-up work, refine their own scope, or retire stories the agent decides are duplicates. The scoped token limits every call to the current Product.

You can't edit or unlink built-in skills.

Custom skills

Custom skills live under your workspace. You write the markdown the same way you'd write any internal documentation — describe the tool, show the invocation, give an example or two. The Agent's prompt appends every linked skill's content underneath the system prompt.

Linking a custom skill to an Agent is done from the Agent's manage skills dialog. Skills are per-Agent, not per-Story: every Story that Agent runs sees the same skill set.

What a Skill is not

A Skill is not code. It cannot exec, it cannot install dependencies, it cannot reach the network on its own. It's prose the model reads. Whether a Skill does anything useful depends entirely on the agent backend interpreting the prose and the tools the backend has available.