When to use

Use directory::prompts::* to surface the static, parametric prompt templates a worker ships alongside its code. Prompts are the slash-command counterpart to the function surface — the user invokes them (/send-email, /triage), and the agent renders them into a real call.

Prompts are sourced from the same skills_folder as skills. Files at //prompts/*.md with YAML frontmatter declaring at least description are exposed; everything else is treated as a skill body. Re-reads happen on every call — file edits are visible immediately, no caching.

The two responses are plain JSON shapes — no MCP envelope, no role/messages wrapper — so this worker stays agnostic to MCP and any other adapter. Adapters can shape the response on their own side.

directory::prompts::list

Inputs

{}

No parameters.

Outputs

{
  "prompts": [
    {
      "name":        "send-email",
      "description": "Compose and send a transactional email.",
      "modified_at": "2026-05-01T12:34:56+00:00"
    }
  ]
}

• name is the prompt's frontmatter name, falling back to the file stem (e.g. send-email.md → send-email). Each name must satisfy [a-z0-9_-]{1,64}.
• description is the frontmatter description (required at scan time — files without it are silently skipped).
• modified_at is the file's mtime as RFC 3339.

Rows are sorted lexicographically by name.

directory::prompts::get

Inputs

{ "name": "send-email" }

name is required and must match the same [a-z0-9_-]{1,64} shape returned by directory::prompts::list.

Outputs

{
  "name":        "send-email",
  "description": "Compose and send a transactional email.",
  "body":        "# /send-email\n\nCompose an email…",
  "modified_at": "2026-05-01T12:34:56+00:00"
}

• name, description, and modified_at mirror the listing row.
• body is the raw markdown body after the YAML frontmatter is stripped — what the user-facing slash command should render.

The shape mirrors directory::skills::get exactly (with name standing in for that surface's id) so a single client struct can target either reader.

Worked example

After directory::skills::download {worker: "resend"} (which writes both the index.md skill body and any prompts/*.md prompt files under /resend/):

{}

→ directory::prompts::list returns one row per prompt the worker shipped (e.g. [{"name": "send-email", "description": "...", "modified_at": "..."}]).

{ "name": "send-email" }

→ directory::prompts::get returns that prompt's body alongside the same name / description / modified_at fields.

Side effects

After every successful directory::skills::download that wrote at least one prompt markdown, the worker fires directory::prompts::on-change with payload { "op": "download", "namespace": "", "source": "repo" | "registry" }. Subscribers (e.g. the mcp worker) use this to forward MCP notifications/prompts/list_changed to their clients without re-polling.

Related

• directory::skills::list / directory::skills::get — same flat shapes for the skill surface (id instead of name).
• directory::skills::download — the only write path. Pulls both skill markdown and prompts into skills_folder from the public registry or a GitHub repo.