Auditing one Claude session: 7.5M tokens, 4,038 tool calls

The author at Kurrent broke down a single 241-turn coding session with Claude: 7.5 million tokens, 4,038 tool calls, and five hours of wall-clock time in a day. The output was accepted, all six plan phases shipped with TDD and review, and the tests scored 5 out of 5.

Where it leaked came from Kurrent’s own tool, Capacitor, which logged the transcript and graded it on a four-axis rubric. Efficiency landed at 3 out of 5, with the turns cited. On turns 7 through 12, Claude confidently named only the PreToolUse and PostToolUse hooks, then surfaced the full lifecycle including Stop once challenged. On turns 130 through 171, a feature got built on an unverified assumption; a 10-minute audit of the real transcripts later showed the needed signal was there all along, and the work had to be redone.

Every “try next time” note from Capacitor is pinned to a specific turn number.