essay · ai-assisted engineering
From "write" to "specify, review, verify"
Every team I know is asking the same question: what does a great engineering team look like when AI writes a lot of the code? Not "will AI replace engineers" — that's the wrong framing. The honest version is harder: what do the humans do when the machines can write faster than we can review?
The answer we landed on at Concirrus, and kept landing on as we rolled it out squad by squad, is a simple loop: specify, review, verify. Humans own the judgement; AI does the writing.
The shift
Before, an engineer's core act was writing code. Interviews tested it, performance reviews measured it, and identity was wrapped around it. AI-assisted engineering inverts that: the writing becomes the cheap part. The expensive parts are deciding what to build, holding the line on how it's built, and proving it actually works.
So we retrained every squad from "write" to "specify, review, verify":
Specify. Acceptance criteria, contracts, and tests come first. If you can't specify it tightly, the AI will confidently build the wrong thing — fast. We treated prompts and specs as engineering artefacts, reviewed like design docs.
Review. CodeRabbit caught mechanical issues; humans reviewed the judgement — security, architecture, business logic. We wrote team norms for review standards, attribution, and what AI-generated code requires before it counts as "done".
Verify. Evals, observability, and production signals became the source of truth. If the system says it works and the numbers agree, it ships.
What changed in hiring
We stopped optimising for typing speed and started optimising for judgement. Hiring now probes: can you decompose a problem into a spec? Can you review a diff you didn't write? Can you tell the difference between "looks right" and "is right"?
Junior engineers got the biggest upgrade. The AI acts as a patient tutor: instant feedback, no ego, infinite patience. Ramp-up time collapsed, and juniors could contribute to production systems earlier — provided the review bar stayed human.
What changed in reviews
Performance reviews now reward the loop, not the output alone: quality of specification, rigour of review, honesty about what was verified. Attribution matters — we track what AI wrote versus what humans shaped, because that's where the real craft is.
The result
100% adoption across squads, close to 3x delivery velocity — with review standards that kept quality and security human-owned.
The machines got faster every month. That's fine. The loop got more valuable every month too — because judgement compounds, and that's the part that stays human.