Dogfooded on our own repositories
The human × AI work timeline, priced
Commits carry a Co-Authored-By: Claude trailer when AI assisted. That alone only
says AI helped. Joining those commits to the actual session transcripts shows
how much human steering went into each unit of shipped work — and once hours
are attributed, every week on the timeline carries a dollar value.
Estimated labour-equivalent value
—
Each band is that week’s range. The headline above is their sum, not a separate calculation.
Every session names the work it did
Topic labels are derived from the transcripts locally. The raw messages never leave the machine that produced them — only counts, durations and these short labels travel.
Method
Accessible, deliberately secondary. Nothing material is hidden behind a toggle.
The rest of the corpus
The same window, read three other ways.
The swimlane
Three lanes. AI-only is zero, and the empty lane is drawn rather than omitted — the absence is the finding.
Labour → compute
Measured tokens at a declared rate card. A model with no rate row is named, not priced at a neighbour’s rate.
The long build and the sprint
Two repositories on one axis, chosen for the contrast in shape.