19 days, 245 sessions with Claude Code: 7 lessons that mattered
A of 9 years spent 3 weeks using Claude Code (86% on Opus 4.8), racking up 245 sessions, 70,000 messages, and 123.1 million tokens. The biggest shift came from using and : writing decisions down once, instead of re-explaining context every session, turns the tool from a smart into something closer to a teammate that remembers past decisions. The real cost isn't itself but repeatedly re-reading the same context in .
To cut that cost, breaking work into phases works well — finishing a task, asking Claude to write a for the next session, copying it, running /clear, and pasting it into a fresh session saved more time than any prompt trick. A meaningful chunk of early manual work also turned out to be tasks a skill or subagent could handle better in a single call.
Key points
- 245 sessions, 70K messages, 123.1M tokens over 3 weeks (86% on Opus 4.8)
- Documenting decisions once in removes the need to re-explain context each session
- The real cost comes from re-reading repeated context, not raw
- Writing a at session end, then pasting it into a fresh session after /clear, beat other prompt tricks
- Many early manual tasks were better handled by a skill or subagent in one call