How I evaluate an AI-first developer
Fundamentals still matter. What separates candidates now is how they direct a coding agent and check its work.
AI-first means the developer stays accountable
An AI-first developer uses coding agents across the whole lifecycle. The agent drafts code, tests and documentation. The developer supplies the context, reviews the output, runs the checks and decides what ships.
My own workflow is AI-exclusive. Claude, Gemini and GPT plan, write, test and debug, and nothing is accepted because a model wrote it. Treat AI-generated code as untrusted until you can explain it, test it and watch it run.
What I require
A candidate missing any of these isn't ready for an AI-exclusive role, however strong the rest of the CV.
- Context
- They give the agent the files, constraints and examples a task needs, and they know when to remove context as well as add it.
- Decomposition
- They split a large problem into tasks an agent can finish and a reviewer can check, and they choose between one model call, one agent and several.
- Verification
- They treat output as untrusted, reproduce failures with logs and tests, and can say which check catches which mistake.
- Tooling
- They use at least one coding agent daily, such as Claude Code, Cursor, Windsurf, Cline, Aider or Copilot. They connect agents to tools through MCP, or a plain API call when that is simpler, and they check a model's tool advice instead of trusting its memory.
- Specification
- They write requirements and acceptance criteria an agent can build against, and they explain progress to people who don't read code.
- Fundamentals
- Data structures, web and backend basics, databases and the full software lifecycle. Without them, a developer can't judge what the agent produced.
Run one real task in the interview
The stack a candidate names matters less than how they work. Give them a small, real bug in an unfamiliar repository and the coding agent they prefer, then watch the session.
Look at what they hand the agent before it starts, how they check its fix and what they do when it is wrong. Then ask them to explain the diff line by line.
A candidate who ships a fix they can't explain has failed the task, even if the tests pass.
What helps after that
Hands-on experience in your stack and your industry shortens the ramp-up. So do test-driven and spec-driven habits, threat modeling for generated code, CI/CD and version control fluency.
The judgment to use one strong agent where several would add only overhead also counts. None of this replaces the requirements above.
Hiring or upskilling for AI-first work?
I work 1:1 with developers and teams moving to an AI-first workflow. Tell me the role or the team, and we can scope the next step.