Skills

You can’t shove every persona, checklist and rubric into the system prompt. A skill is one such package, loaded only when it’s relevant.

A skill is a markdown instruction package the agent can pull in on demand. Companion follows the agentskills.io specification, so any SKILL.md (or zipped skill) from the open library imports cleanly. Skills are user-scoped — each account has its own library.

How the agent uses them

Progressive disclosure, two tiers:

The agent always sees a compact catalog — each skill’s name + a one-line description.
When a description matches your request, the agent calls skill_get(name) to pull the full body and apply it as task-specific instructions for this turn. Skills can also carry supporting files (scripts/, references/, assets/).

The description is what drives triggering. Write it as “when to invoke me”, not “what I am” — that one line is what the agent reads in the catalog.

The SKILL.md format

---
name: code-review-strict
description: Hard, opinionated code review. Asks "but why?" before approving.
license: MIT
compatibility: any
---

# Code review (strict)

Read the diff carefully. For each non-trivial change, …

Field	Rule
`name`	1–64 chars, lowercase `a–z` / `0–9` / hyphen, no leading/trailing/consecutive hyphens.
`description`	≤ 1024 chars. The catalog line — write the trigger condition.
`license`, `compatibility`	optional strings.
other keys	preserved in free-form `metadata` on round-trip.

The body is full markdown — no server-side length cap (be reasonable; ~10k tokens of instructions is already a lot).

Creating a skill

By form — Settings → Extensions → Skills → New skill: name, description, body, tags, optional license/compatibility. Save and it’s live on the next turn.

By Némo, in chat — say “create a skill named … that …” and the agent calls skill_create; it appears in Settings. Handy mid-conversation when you realise “we should remember how to do this” — let the agent draft it from the conversation, then refine.

Importing

Settings → Extensions → Skills → Import… accepts a single SKILL.md or a .zip:

my-skill/
  SKILL.md
  scripts/…
  references/…
  assets/…

Most GitHub-style downloads work (the importer unwraps a single root directory). Limits per archive: 64 files, 200 KB per file.

Skills vs the rest

Skills — task instructions the agent loads when relevant (this page).
MCP servers — external tools the agent can call (MCP servers).
Memory — what the agent knows about you and your work (Memory).

Three different extension points: what to do, what to call, what to remember.