Week of 2026-04-27
This report is private to you. On auth-svc: shared with the project lead (set at project-join). On billing: not shared (your default). Change anytime in /settings.
Worked on
Project-by-project breakdown of where this week's sessions landed.
credential rotation work
test fixture cleanup
invoice retry logic
smaller one-offs
Token usage & cost
A cross-session rollup of what your Claude work cost this week, broken down by model with cache-hit and output:input efficiency.
Costs are estimated from per-model list prices. Cache reads are billed at a fraction of fresh input — a high cache-hit rate is most of why this number stays low.
What's working & what to watch
Patterns Claude noticed across this week's sessions. Surfaced as observations, not as a grade.
Scoped, hypothesis-driven prompts on the credential work ("here's the error, here's what I think is happening, verify before changing code") — these one-shot consistently.
Three sessions started with a paste of the full stack trace and no context. Claude's first reply was usually wrong; you corrected on turn 2. Worth front-loading the hypothesis the way you do on the credential work.
Lessons that match what you did this week
Pulled from the team library. Read these — they're load-bearing for the patterns above.
Front-load the hypothesis, not the trace
Pasting a stack trace alone gives Claude permission to guess. Stating your hypothesis first turns the turn into verification, which is cheaper to be wrong about.
Scope rm operations explicitly — never trust 'old' as a filter
When asking Claude to clean up files, always scope by an objective signal (unreferenced in code, last-modified > N days, matched glob). 'Old' is interpreted aggressively.