Capstone
Every module drilled one dimension in isolation. They trade against each other, and the skill is holding all of them at once under a real budget.
Lesson 1
The dimensions fight each other
Up to here each module let you optimise one thing. In practice they pull in opposite directions, and the interesting decisions are all trades.
Specificity costs tokens; the budget wants them back. Structure costs tokens too, and earns them back the third time you edit the prompt. Examples are more effective than rules and more expensive per idea. Delegating arithmetic to a tool costs a round trip and buys you the elimination of a whole error class.
There is no setting where all six are maximised. What a competent practitioner produces is a defensible allocation — and the tell is that they can say what they traded away and why.
Lesson 2
The order that saves the most time
Start minimal. Take the strongest model, give it the least you can, and find where it actually fails. Add instruction only against observed failures.
This is the opposite of how most prompts get written, which is to anticipate everything on the first draft. Anticipated failures are cheap to imagine and expensive to carry: most of them never occur, and each one costs budget on every call forever.
The failures you observe are real by construction. Patching only those converges on a prompt that is minimal without anyone having to guess what minimal was.