Lesson 4 · Part 1
From “it can” to “it does, reliably”
Ask any AI to plan a party and it will produce something. Ask it twice and you'll get two different somethings. A skill is how you fix that: a written recipe for a task you want done the same way every time — with a name, a trigger, steps, and a way to check the result.
Toodles has one called plan-clubhouse-show. The trigger: Mickey asks to plan a show, or the morning's problem arrives from the Mousekedoer. Then five steps, including “match each segment to a friend's strength,” “reserve one Mousekatool for the finale,” and — step four, immovable — “schedule the Hot Dog Dance as the closer. Always.”
Lesson 4 · Part 2
The test plan is the grown-up part
Here's what separates a real skill from a hopeful prompt: every skill ends with a test plan — plain sentences describing what “it worked” looks like. For the show planner: every segment has a named owner; the Hot Dog Dance appears as the final item; no segment requires a tool the Mousekedoer didn't provide.
That list means you can actually check the assistant's work — and in advanced setups, the checking itself gets automated: the test plan becomes a real test the AI is measured against. “Looks reasonable” is not a quality bar. “Passes its test plan” is.
Lesson 4 · Part 3
A library, not a lecture
Skills accumulate one file at a time: plan-clubhouse-show, pick-mousekatool, cheer-up-donald. Each is small, named, and independently improvable — fix the recipe once and it's fixed everywhere, forever. Compare that to re-explaining what you want in every chat, from scratch, for the rest of your life.
The takeaway: if you'll ever want a task done twice, write it down once — trigger, steps, and how to check it worked.