Session 10. Skills and an architecture decision record — Fri 25 Sep
Conclusion
You wrote a skill: five sections, in a file, loaded when its description matches the task instead of retyped every time.
What you did
- The
descriptionis routing and the body is the instructions, which is why ten skills cost you ten lines of context until one of them is needed. - The failure rules are the part that changes an output; the safety boundary is a sentence, and the thing that actually stops a write is still the tool contract.
- You proved the skill did something the only way it can be proved: the same task
run without it and with it, two excerpts that differ, and one instruction you
fixed after watching it fail.
ch10-e1refuses two identical runs, because two identical runs are the evidence that nothing happened. - Then you ran one assistant two ways — the framework-free loop and the same four nodes as a declared graph — and counted the model calls in each.
- You recorded one architecture decision you actually made: what you chose, the option you turned down, why it lost today, and the number that reverses it.
The failure you handled
A framework used where a fixed workflow would have been better. Known steps, an unvarying sequence, and a planner bought to make a decision the code already knew how to make.
What to carry forward
ch10-e2 refuses "when it gets slow" and accepts "p95 over 2000 ms for 15
minutes", the same standard the depth track's architecture module uses, because a
decision nobody can falsify is a preference.
Into session 11
Session 11 assumes both artifacts exist and starts week 3 on state and memory, where the storage policy gets the same treatment: written down, with something it refuses to remember.
Weekly challenge 2
Your store, and a buyer that refuses well. Due Monday. It is step 1 of your final project, and its score (up to 500) is added to this session's. Write it in the challenge cells at the end of the session notebook.
Homework
Swap both artifacts with another learner. Review their skill for ambiguity and unbounded permissions, then ask whether you could tell, this week, if their trigger had fired.