AI Pair Programming Isn't Autocomplete: A Real Workflow With Claude Code + ChatGPT
The actual day-to-day mechanics of working with two different AI tools on the same codebase — not a highlight reel, the boring parts that make it work.
"AI pair programming" gets described like autocomplete with better guesses. In practice, on a real solo project, it looks more like having two very fast, very literal collaborators who need clear instructions and will confidently do the wrong thing if the instructions are vague.
The actual division of labor
- Claude Code, in the terminal, for anything that touches the real codebase — multi-file changes, running the test suite, debugging against actual stack traces and logs, where having direct tool access to the repo matters.
- ChatGPT for scoping a problem before committing to an approach — thinking through tradeoffs, sanity-checking an unfamiliar API, drafting the shape of a solution I'll then implement deliberately rather than paste in wholesale.
- GitHub Copilot for the small, local, obviously-correct completions inside a file I'm already actively editing — the least interesting use, and still a real time saver on repetitive patterns.
The discipline part nobody mentions
Every AI-generated change gets read before it's accepted, the same as a human teammate's PR would. The failure mode isn't the tool being wrong occasionally — it's a developer who stops reading because the tool is usually right. "Usually" is exactly the gap where a subtle bug survives to production.
The workflow that holds up: describe the outcome precisely, let the tool produce a first pass, then review it as if a junior developer wrote it — because functionally, that's what just happened.
A car rental agency needed to move off manual, phone-and-spreadsheet bookings to a real-time online platform — built and run solo.