SKILLS, RECIPES, ROUTINES
// the HOW, the WHAT, and the WHEN of repeatable work
The three-layer model
| Layer | Answers | Example |
|---|---|---|
| Skill | HOW to do something well | "how we write release notes" |
| Recipe | WHAT to produce | "weekly competitor digest" |
| Routine | WHEN it happens | "every Monday, 9am" |
Skills teach technique, recipes define deliverables built from those techniques, and routines put recipes on a schedule. Each layer is inspectable and editable — no black-box "automations."
Skills
A skill is a packaged instruction set an agent loads when the task calls for it. Agents ship with class-appropriate skills, and gain more as you work together. Skills are capability-gated: a skill that needs a tool the agent hasn't been granted stays visibly dormant instead of silently failing — the skills panel always distinguishes what you chose to grant from what the station floor provides.
Recipes
A recipe captures a deliverable you'll want again: its inputs, its steps, its output format. Run one on demand, hand it to any capable agent, or attach it to a routine. Your usual recipe is offered as a one-tap prefill when you return to a session.
Routines
Routines are the schedule layer — see Autonomy & Night Shift for how scheduled runs execute and report. A routine is always a real recipe run by a real agent; there is no separate, weaker "automation mode."
The loop closes: work that sharpens itself
Rate a delivered run and the station extracts why it landed (or didn't) into your taste profile. Recipes carry those learnings forward, the quest board regenerates around your current goals instead of stale ones, and the scout studies your history to pitch next moves you didn't think to ask for. A routine left running doesn't decay into noise — each cycle is built with more evidence of what you value than the last.