Five questions to answer in fifteen minutes. The numbers below are pulled from your traces. The point is not to optimise — it's to pick one specific change you'll make next week, and check next Sunday whether you actually made it.
Why — usually one of: an underspecified question, a session that went in circles, or too much bundled into one prompt. Name the cause for each in one sentence.
Patterns with a low cache rate are usually prompts whose preamble varies session-to-session. The fix is a template.
Whatever it is, that template is the one most worth tightening — the saving compounds.
max_tokens / refusal / pause_turn stop reasons, plus any session with status error. These are the failures you'd otherwise scroll past.
Concrete, not aspirational. Not “be better at prompting” — something like “use plan-mode whenever I'm touching three or more files” or “never start a new task in the same Claude Code session as an old one”.