---
title: Agent Skill generation
description: A Skill pack from your spec, carrying your own prose with it.
---

`reed skill generate` produces an Agent Skill pack — a `SKILL.md` and its supporting
files — from a configured OpenAPI spec.

```json
{
  "skill": {
    "ingestion": {
      "output": "generated/skill-ingestion",
      "name": "acme-ingestion",
      "concepts": ["concepts/events", "concepts/identity"]
    }
  }
}
```

Then:

```sh
reed skill generate
```

<ParamField name="output" type="path" required>
  Where the pack is written, inside your repository.
</ParamField>

<ParamField name="name" type="string">
  1–64 characters, matching `^[a-z0-9]+(-[a-z0-9]+)*$` — the Agent Skills format's own
  constraint, not one Reed invented.
</ParamField>

<ParamField name="concepts" type="string[]">
  Page references bundled verbatim into the pack.
</ParamField>

## `concepts` is the part that matters

A spec describes endpoints. It does not describe what your concepts mean, which call order
is correct, or which failure is retryable — and those are exactly what an agent gets wrong
without help.

`concepts` names pages of your own documentation to bundle into the pack verbatim. The
prose you already wrote for humans goes with the tools.

References are validated twice: the schema catches malformed entries and duplicates within
the array, and a content-aware check catches a reference that resolves to no real page. An
unresolvable concept fails the run rather than silently bundling nothing — a Skill missing
the one page that explained the call ordering is worse than a Skill that refused to build.

## It lands in your repository

Like MCP generation, `output` is a path in your project. Reed does not host or distribute
the pack, and generating it changes nothing about your rendered site.

## Keeping it current

Regenerate when the spec or the bundled pages change, and commit the result. Because the
concepts are references rather than copies in your config, editing the underlying page and
regenerating is enough — there is no second copy of the prose to keep in sync.