30 Days of AI Collaboration

The shape of April

Daily session volume across the month. The shape reflects real work rhythm: 04-28's 30 sessions (1.9× the daily average) aligns with launch-week density on Silo iOS. The April 30 spike is a Gemini ingestion batch with shorter sessions, not a real day-of-work; the Apr 22 dip is a known travel day.

Tools and projects · friction signature

Friction-heavy rate by tool and by project. Claude Code is the dominant client; Gemini CLI runs at much lower friction (shorter sessions, lighter scope). The project view shows where the heat lived: Silo absorbed the most friction-heavy sessions (81 of 135), driven by iOS launch-week complexity.

Session length · the long tail absorbs the time

Distribution of sessions by turn count. Most sessions are short and tractable. The rose-colored long tail is the friction zone: 140 sessions (30.0%) ran past 100 turns and absorbed 86.3% of total dialogue. The audit's correlation between high turn count and architectural drift lives in those bars.

Token economics · partial coverage

Token accounting was wired into Claude Code mid-month, so this view is partial: 225 of 466 sessions have token data, and the patterns from those don't generalize to the whole corpus. Useful as a baseline order-of-magnitude.

The compounding curve

Cumulative decisions (orange) and sessions (cyan) across April. The curve is roughly linear but accelerates after April 9, the inflection point where launch-readiness work began stacking on top of normal daily output. Decisions accumulated faster than sessions did, which is the compounding effect: the Vault as input made each successive session more decision-dense.

Recommendation → decision conversion rate

When the AI made a recommendation, how often did it become a recorded decision in the same session? thedetech ran at 128.1% while Finance sat at 36.0%. The first means I was reaching decisions faster than the AI recommended them; the second means recommendations stacked up without commitment. Conversion can be misleading. A rate over 100% means more decisions were recorded than recommendations issued, the human was leading.

Open questions atlas · the knowledge boundary

868 open questions across April, classified by theme. The largest categories are where AI consistently fails to close the loop: strategy/roadmap calls that need taste, architectural decisions that need codebase context, and debugging questions that require live system access the model doesn't have. These are the human's job, still.

Top 12 technologies by session-mention

Granular tool mentions, deduplicated within each session. Git tops the list at 48.7% of sessions with tech metadata, nearly half the work touched version control in some way. The Rails-Swift dual gravity is real and almost evenly split. The local-AI stack (Gemini CLI, LM Studio, Ollama, Claude Code) collectively shows up across more sessions than any single language.