procedural (memory, project_template)

Procedural memory

Format

One file per skill, named {slug}.md:

---
name: slug-matching-filename
trigger_when: human-readable trigger condition (when should the agent reach for this?)
---

## Steps
1. ...
2. ...

## Gotchas
- ...

## Last validated
YYYY-MM-DD

When to write

A skill is written only when a workflow has succeeded at least twice. One success is not a pattern — it's an episode. The session-closer enforces this rule.

Writing a skill from a single success creates fabricated procedural knowledge — the most dangerous kind of memory pollution. The skill-finder will trust whatever is here.

How a skill graduates from learned to user-built

Skills synthesized by the agent live in .xo/skills/learned/. If the user reviews and endorses one, they may move it to .xo/skills/user-built/ — those are treated as authoritative.

Editing

Skills can be edited, unlike episodes. When you re-validate a skill on a fresh success, update the Last validated field. If a step changes, edit it and note the change in the most recent session's session.md.