# Workflows — Jarvis's teachable routines

One file = one routine Ahmed taught. When he says its trigger phrase
("gaming time"), Jarvis reads the file and executes the steps top-to-bottom,
fast, with **no questions** — the teaching already happened.

## Format (markdown, one file per workflow)

```markdown
---
name: gaming time
triggers: gaming time, game time, let's game
---
1. launch Steam (mcp__desktop__launch "steam")
2. wait ~3s for the window
3. maximize Steam (window.json {"action":"fill","apps":["Steam"]})
4. key ctrl+1 (library tab)
5. click 480,220   # "Play" button — coords recorded with Steam MAXIMIZED
...
```

## Rules Jarvis follows when saving one
- Save to `workflows/<slug>.md` with the trigger phrases in the header.
- Prefer **shortcuts / launch / CLI / deep links** for steps; use raw
  coordinates only when nothing else works.
- Any coordinate step MUST note the window state it was recorded in
  (maximized / fullscreen / windowed WxH at X,Y) — and the workflow should
  put the window in that same state (fill/maximize it) BEFORE clicking, so
  the coordinates stay valid forever.
- After teaching, confirm back the saved steps in one short sentence.