A workflow for AI pair programming — delegate, discuss, do it yourself
The developers getting the most out of these tools are not the ones using them the most. They are the ones who have a rule for which of three modes a task belongs in, and apply it without thinking.
AI Tools Hub Editorial · published · updated
The three tiers
| Tier | What goes here | Review needed |
|---|---|---|
| Delegate | Boilerplate, tests for existing code, mechanical refactors, format conversions | Read the diff |
| Discuss | Approach for an unfamiliar API, debugging a stack trace, naming, trade-offs | You write the code |
| Do it yourself | Security-critical paths, architecture, anything with institutional context | No AI output in the commit |
Delegate
The test for this tier is whether success is mechanically verifiable. Tests pass or they do not; the conversion is correct or it is not. Hand it over entirely, then read the diff as you would a colleague's pull request — because that is what it is.
Discuss
This is where most of the value is, and where the usage statistics do not show it. Asking why an API behaves a certain way, what the options are, what breaks under load — then writing the code yourself. You keep the understanding, which is the thing you will need in six months.
Do it yourself
Authentication, permissions, payment, anything load-bearing for security. Also anything shaped by context the model cannot see — the module being deprecated, the pattern that caused an incident. Generated code here is syntactically correct and organisationally wrong.
Making it a habit
- Decide the tier before you start, not while you are stuck
- If you are in Delegate and reviewing takes longer than writing would have, it was the wrong tier
- If you are in Do it yourself and the task turns out mechanical, move it down
- Write the tiers somewhere the team can see them, so reviews can refer to them
Does this slow you down?
The categorisation takes seconds and saves the much longer loop of generating something, reviewing it carefully, and discarding it.
How does this work for a junior developer?
Shift the boundary: more Discuss, less Delegate. The understanding is the point of the first few years, and delegating past it is expensive later.
Related
More articles
Opinion
Why AI rollouts fail in Japanese companies — four structural reasons
For business
Enterprise AI rollout checklist — SOC 2, ISO 27001, DPAs and personal data
For business
Japanese-developed LLMs for business — ELYZA, Sakana AI, NTT tsuzumi and PLaMo
Comparison
AI coding tools compared — Cursor, Copilot, Claude Code, Windsurf and Codeium