About

Summary of what my Claude Code usage analytics say about how I work (April–May 2026):

241 sessions · ~109K lines across 1,407 files · 88% of classified outcomes mostly-or-fully achieved.

• §Design first, code second. Most features begin as a structured interview with Claude Code grounded in the existing codebase, locking decisions one-by-one until there is a persistent design brief and a phased plan. Implementation starts only after that.
• §Phased plans, TDD per packet. Work decomposes into named packets (P0.A → P5.E) and executes under strict red-green-refactor cycles. Of 130 classified outcomes, only 7 hit “misunderstood request” — the planning happens before the prompt.

• 

• §Multi-Claude orchestration. An orchestrator session dispatches packets to worker sessions in worktrees, each verified at the diff before merge. ~25% of messages cross between them.
• §Vertical-slice ownership. A single session routinely threads schema migrations → typed backend → React + Zustand frontend → Raspberry Pi runtime. Multi-file change is the #1 capability the agent contributes.
• §Active supervision. Reads run roughly 1:1 with edits: agent output gets checked, not assumed. Every “wrong approach” or “buggy code” event (32 of each over the month) is caught at the diff, not in production.

Agent as execution surface. Architecture, module boundaries, and data contracts get worked out earlier on a separate surface (usually Claude Web with the project context loaded), so Claude Code starts from decisions already made.