10.2 billion tokens across 520,000 messages in 2,346 sessions. That's what one year of heavy Claude Code usage looks like, according to developer Max Ghenis, who published a detailed breakdown of his setup and spending.
The cost trajectory tells a story on its own. At peak usage in August 2025, Ghenis was spending $5,861 per month on API calls. His biggest single day hit $505. He's since brought that down to roughly $200/month by switching to Anthropic's Max plan subscription, which bundles Claude Code access at a flat rate.
The Configuration That Survived a Year
Ghenis built out 12 custom slash commands, five automated hooks, and a skills system that loads domain knowledge on demand. The practical highlights:
/briefingpulls together his calendar, emails, and task reminders into one view/search-everythingsearches across local files, WhatsApp, Gmail, meeting notes, and browser tabs simultaneously/bouncesends implementation plans to GPT for a second opinion before he commits to them- An enforce-package-managers hook that blocks
npmandpipcommands, forcingbunanduvinstead. His take: putting this in a hook instead of CLAUDE.md instructions changed compliance from "maybe half the time" to 100%.
That last point is worth stealing. Hooks run as shell scripts before or after specific Claude Code actions. They enforce rules mechanically rather than hoping the model follows written instructions. If you have a rule Claude Code keeps ignoring, a hook is probably the fix.
What Actually Worked vs. What Didn't
The most honest part of the writeup is the failed iterations. Ghenis went through three terminal setups (VS Code with extensions, custom grid layouts, AppleScript automation) before landing on a simple iTerm2 plus tmux configuration that required just three small config changes instead of hundreds of lines of custom code.
He also discovered 36GB of stale repository clones and a plugin that had silently hijacked his git history, problems he only found by deliberately auditing his setup.
His global CLAUDE.md instructions file shrank from 113 lines to 70 after he moved domain-specific knowledge into 12 separate "skill" files that load only when relevant keywords come up in conversation. This keeps context windows clean instead of front-loading every session with information that's rarely needed.
For anyone running Claude Code as a daily driver, the configuration repo is public on GitHub. The biggest takeaway isn't any single trick. It's that the tool rewards deliberate, ongoing maintenance of your setup the same way a good IDE config does.