Claude Code in 12+ sessions: grep xcodebuild test to cut tokens

The central mechanic the developer exploits is simple: Claude Code resends the full conversation every exchange, so the cost of a session is not the prompt alone but everything previously added to it. That is why the biggest win comes from keeping long build logs out entirely—grep-ing xcodebuild output down to error and result lines means a suite of 400-plus tests run dozens of times still adds only a few lines to the conversation. The same logic drives the other techniques: contact sheets of thumbnails instead of many full screenshots, grep -n plus sed to read only a range of a 2,000- or 9,000-line file, and checking page state through JavaScript text output rather than visual inspection. The developer also describes packing multiple browser or shell actions into a single call, since each round-trip re-sends the whole conversation.

Running a dozen-plus sessions makes these small savings compound, and it also creates new failure modes. Summarization is one: when a session gets long, Claude Code summarizes it automatically, and if detailed rules drop out of the summary the agent has to re-research them, adding cost. The developer's answer is to keep agreed rules, requests, drafts and handoffs outside the conversation—in memory files, GitHub Issues, docs/ and HANDOFF.md—so only diffs and key points are pasted in. Another failure was a hook that shortened command output interfering with auto-approval, blocking even harmless counting commands.

The developer reports that token-saving has costs as well as benefits: trimming output too aggressively once hid a test failure, and a broken version was committed, so tests are now run separately before committing. Language drift after summarization—replies switching to English mid-session—was addressed by writing an explicit 'no exceptions, Japanese' instruction at the top of CLAUDE.md.