Skills

A skill is a page in the skills/ directory that packages reusable instructions: a house tone, a checklist, a domain playbook. An agent discovers its bound skills by name at run start and loads a skill’s full body only when it decides the skill is relevant. This keeps the system prompt small while giving the agent a library it can reach into, and it mirrors the Agent Skills package format so one library serves Subspace agents and external harnesses alike.

Skill pages

Skills live in skills/ and carry a metadata.skill block:

---
skill:
  name: pitch-tone
  description: A concise, confident sales tone for outbound follow-ups.
---

# pitch-tone

Lead with the prospect's stated problem. One idea per sentence. No hedging.
Close with a single, specific next step. Never invent a discount or a date.

name defaults to the page title when omitted; description is the one-line summary used for discovery. The page body is the skill itself.

Binding and discovery

A definition binds skills with skills: string[] (by name or slug); absent means all skills are bound. At run start, the runtime injects one discovery line per bound skill into the run’s system message, next to the memory block: each line pairs the skill’s name with its description, so the model knows a skill exists and what it is for without spending tokens on its full body.

Note

Discovery is name and description only. The agent sees that pitch-tone exists and is “a concise, confident sales tone for outbound follow-ups”, then chooses whether to load it. A definition bound to twenty skills injects twenty short lines, not twenty full pages.

Loading a skill

When the agent decides a skill is relevant, it calls the ungated skill.load {name} tool, which returns that skill page’s body rendered through the OKF codec. Loading is ungated because reading an instruction page is not an outbound action.

Caution

Loading a skill never expands privileges. Whatever a skill instructs still routes through the tool registry, the definition’s tool allowlist, and the approval gates: the alwaysGate trust floor holds. A skill that says “send the email” still pauses for your approval, and a skill cannot call a tool the definition did not allow.

SKILL.md package interop

On disk, the OKF bridge maps each skills/ page to a standard Agent Skills package directory: skills/<slug>/SKILL.md with name and description frontmatter, and any attached files written alongside. Import reads the same layout back into pages. So a skill you author in the outliner exports as a portable SKILL.md package, and a SKILL.md package dropped into the OKF folder imports as a skills/ page.

okf/
  skills/
    pitch-tone/
      SKILL.md          # name/description frontmatter + body
      examples.md       # attached file, preserved on round-trip

This is why one skills library can serve both Subspace’s own agents and external agent harnesses that read the Agent Skills format: the source of truth is a page, and the file tree is a faithful, editable mirror.

Related

Agent definitionsHow a definition's skills: list binds skills to a run.

OKF bridgeThe codec and the export/import round-trip for skill packages.