Claude Code in production: a workflow that actually ships
Claude Code earns its keep when you treat it as an engineering pair, not a snippet generator. Hooks, MCP servers, and a tight CLAUDE.md are the difference between toy demos and shipping code.
I've used Claude Code daily for eight months across this site, a couple of commercetools surfaces, and a Mars-side service. The first month was magic. The second was disappointment. The eight-month version is steadier and more honest about where the tool earns its keep.
What works: a tight CLAUDE.md that captures the project's conventions in 200 lines or less, a couple of MCP servers wired up for the boring lookups (Linear, the docs site, an internal API explorer), and slash commands for the workflows you actually repeat. The settings.json hook system is the underrated piece, blocking destructive shell commands and auto-formatting on stop turn the agent from theoretical risk into something you trust on a real branch.
What doesn't: asking it to drive a refactor without a clear boundary. The model is good at local edits and bad at understanding architecture you haven't explicitly mapped. The fix is the same as for human collaborators: write the plan first, get alignment, then implement. Skipping that step is how you end up with three layers of speculative abstraction in a PR.
The biggest unlock this year was subagents. Spinning up an Explore agent for a search task while the main thread keeps building is genuinely a productivity multiplier. The wrong move is to delegate understanding; the right move is to delegate lookups, code-grep work, and tedious enumeration. Treat them like junior engineers with a clear question.