BlogDevelopment
Shipping with AI without losing control
Coding agents as copilots: reviews, boundaries, and flows where humans own architecture and merges.
AI doesn’t replace judgment
I use Cursor, Claude Code, and Warp daily to write, debug, and document faster. What I don’t delegate: architecture, security, and “should this even exist?”
Practices that work for me
- Short, verified context — real file/error, not novels
- Small changes — reviewable PRs, not monsters
- Tests as a net — Playwright / unit where it hurts
- Secrets out of chat — env vars, never keys in prompts
Client outcome
Less time on boilerplate, more time integrating systems (CRM, APIs, bots) where business value is obvious.
The useful question isn’t “do we use AI?” — it’s “who owns the final decision?”.