A cache miss costs you money, never memory
Prompt caching does not hold your conversation, so it cannot lose it. Unreliable savings, perfectly reliable correctness.
These are about one question: how do you know when you have practiced enough? Every post here argues from the published blueprints or from measurements on our own bank. None of them tells you what an exam costs or how long it runs — Anthropic publishes that, and repeating it badly is how a study site loses the reader it just earned.
Prompt caching does not hold your conversation, so it cannot lose it. Unreliable savings, perfectly reliable correctness.
Most agent designs go wrong by imposing a structure the work does not have. Draw the dependencies and the pattern names itself.
Whatever the brief leaves out, the subagent has to guess.
A plan to approve, a gate, a retry limit, a customer asking for a human. Each is a pause, and each belongs somewhere different.
Tool descriptions, resource lists and error messages all end up in front of the model. When Claude picks the wrong tool, the fix is usually in text the server wrote.
The usual mistake with Claude Code configuration is a good instruction filed in the wrong place. Two questions tell you where it belongs.
Any tool that re-asks the questions you got wrong poisons its own accuracy statistic. Here is how far off it gets, measured.
Scaled scores and weighted blueprints both break the arithmetic people do in their heads.
Being terrible at something worth 2.6% of the exam is cheaper than being mediocre at something worth 33%.
A single number with no interval around it is a guess wearing a lab coat.
Everything above is the reasoning behind one number. The diagnostic produces yours in twenty questions, with the interval that says how much to trust it.