Session 10. Skills and an architecture decision record — Fri 25 Sep

Quick quiz (ungraded)

So the file is easier for a human to skim
A side effect. The reason is about what the model loads, not about reading.
The description is always in context and the body loads only when the task matches, so ten skills cost ten lines until one is needed
Correct. That is progressive disclosure, and it is context budgeting.
Because the assistant executes the description and ignores the body
Nothing in a skill executes. The model reads both; the difference is when.

The skill needs stronger wording
Stronger wording is still wording. The next model, or the next long context, ignores it again.
Nothing enforced it — a safety boundary in a skill is a hint, and the tool that could edit was still available
Correct. Permissions live in the tool contract and the assistant's configuration; the skill says what should happen, not what can.
Skills cannot express safety boundaries, so the section should be dropped
Write it — it changes behaviour often. Just do not confuse it with a permission system.

Identical runs prove the skill is stable
They prove the skill changed nothing. Stability of an output you did not influence is not evidence about your file.
The skill made no difference, so there is nothing yet to keep or improve
Correct. A skill is justified by the difference it makes; no difference means delete it or sharpen it.
The check cannot compare two different strings
It compares them fine. The refusal is a judgement about the evidence, not a limitation.

None, as long as the output is correct
The output is correct until the day one hop fails. You added failure modes to buy a decision that was never in doubt.
Model calls and failure modes bought a decision the code already knew how to make
Correct. When the steps are known in advance, a fixed chain is the design: fewer failure modes, no spend on deciding, a predictable trace.
Only the extra dependency
The dependency is the cheapest part. The hops, the spend, and the unpredictable trace are the rest.

"When the retriever starts to feel slow for users"
No number. Two people can hold this belief and disagree about whether it has happened, forever.
"When we go past 2000 and it hurts"
A number with no unit. 2000 what? Nothing can be compared against it.
"When p95 for one answer stays over 2000 ms for 15 minutes"
Correct. A number and a unit means somebody can check on a Tuesday and get one answer.

So the reader has an alternative to try if the decision fails
Useful, but the check is not about a fallback plan.
One option is a justification written after the fact; two is the smallest record of an actual choice
Correct. A record that never mentions anything you did not do is defending the only thing you tried.
Because every architecture question has exactly two answers
Most have several. Two is the minimum the check can insist on, not the true number.