# Session Pattern Analysis — February 2026 Analysis of 682 Claude Code sessions across 40 days (Jan 5 – Feb 14, 2026), drawn from the session index. Data, not diagnosis. --- ## The Numbers 682 sessions in 40 days. Every day, no gaps. Average ~17 sessions per day. 59% of sessions are 5 turns or fewer. Research sessions (generating philosophy, typography, propaganda, etc.) account for about 12.5% of total turns. System/config/vault architecture accounts for 16.6%. Routine (morning review, calendar, tasks, email) accounts for 11.8%. And 52% of all turns are in sessions that don't match any project keyword — a mix of vault file operations, tool troubleshooting, content processing, session recall, editor-context work, and miscellaneous. Research sessions are the deepest engagements when they happen — 11.6 turns average, highest of any category. But they're only 9.4% of sessions by count. Time-of-day is bimodal: a broad plateau from 11:00–14:00, then a second surge at 22:00–23:00 (over 100 sessions in those two hours combined). --- ## Three Things You Might Not Be Seeing ### 1. The generating philosophy project has a specific false-start fingerprint Feb 10: four sessions with the same opening ("Explore sources for generating philosophy paper") — turn counts of 3, 16, 4, 50. Three attempts before a productive session. Feb 11: four sessions with "I'm a bit stuck" — 2, 1, 2, 2 turns each — before other sessions got traction. Feb 12: five "continued from" sessions (the opposite problem — productive work that kept hitting the context wall). This isn't a general pattern across all your work. It's specific to this project, and it's concentrated in the period when you were working on the hardest parts (source reading, Section 1 rewrite, intro). The pattern is: difficulty getting into the session → once in, difficulty being forced out of it by context limits. The friction is at the threshold, not in the work itself. Once a session passes ~5 turns on this project, it tends to run long. I don't know whether the false starts are technical (PDF crashes, context loading failures) or something else (uncertainty about how to proceed, the session opening going in a direction that doesn't feel right). The data can't distinguish those. ### 2. The "other" bucket is the majority of your Claude time 52% of turns are in sessions that aren't about any named project, system design, routine task, or personal matter. They're the operational fabric: finding files, processing content, figuring out tools, recalling past sessions, responding to editor context. This is probably invisible to your self-model of your workday because each individual instance feels trivial. But collectively, it's where most of the Claude interaction goes. This isn't necessarily a problem. It might just be what using a computer looks like. But if you ever feel like you've been at it all day and haven't produced much on the paper — this is probably where the time went. The research share of Claude time is smaller than the research share of your *attention* feels like, because the connective tissue is extensive. ### 3. System work and research work are roughly equal in volume, and both are intensifying System/infrastructure: 16.6% of total turns. Explicit research: 12.5%. (The true research share is probably higher because the keyword filter misses some research-adjacent sessions in the "other" bucket.) The most recent week has the highest levels of both simultaneously — 258 research turns and 236 system turns. I'm stating this as a fact, not a diagnosis. You already have guardrails in the config against over-building ("don't suggest elaborate reorganization"). Whether the current ratio is right is a judgment only you can make. The data just says: the two activities are comparable in scale and both growing. --- ## What I Can't See The metadata tells me when, how long, and roughly what topic. It doesn't tell me: - Whether short sessions were productive (a 1-turn session that files a PDF successfully is fine) - The quality of thinking in any given session - What happens between sessions — whether ideas develop in the gaps - Whether the late-night work is your best or most scattered - Whether the "other" bucket is friction or just normal computer use - Whether the false starts on the generating philosophy project are technical failures or something about the project itself The strongest claim I can make from the data alone: the generating philosophy project has a distinctive start-up friction that other work doesn't show, and the majority of your Claude interactions serve the connective tissue of the workday rather than any single project. Everything beyond that is interpretation.