Bad ideas that aren't bad anymore
A veteran engineering leader lists practices he used to reject, like skipping human PR review, splitting teams across separate work streams, and going shallow instead of deep, that now work well because AI tools like Claude have cut the cost of context switching and code review. He admits his gut instincts may be less reliable than they were.
- Code review was replaced for about a month by having Claude author PRs and the ticket owner review them, with Greptile and CI as additional checks. The ticket owner understands the problem best, and the tools catch small bugs well, though big architectural issues are still rare.
- Splitting the team into parallel work streams no longer creates knowledge silos, because a new person can ask Claude to summarize a colleague's work and get up to speed in minutes. The real skill is now grasping the problem and solution shape rather than the code.
- Deep specialization in one framework matters less, since AI can answer framework questions instantly. Strong engineering fundamentals like abstraction design and judgment, plus product, design, and user research skills, are now the baseline for value.
A veteran engineering leader with about 15 years in software says that some of the intuitions he's relied on have started failing, and he's testing that by changing how his team works. The first change: about a month ago his team stopped doing human code review. An engineer works with Claude, Claude opens the PR, and the engineer who owns the ticket reviews it. Once CI passes and the Greptile review comes back clean, that same engineer merges if they feel confident. His argument is that the second set of eyes is still there, since the human owner understands the problem better than a teammate paging in from scratch, and Claude and Greptile are good at catching small bugs and inconsistencies. Architectural flaws and unintended product impact are rare at this stage, he says, so the review gate was mostly guarding against the wrong things. He admits it's only one month in.
The second change came out of a retro. With several people on holiday, the team split into separate parallel streams, each person owning one. He had always resisted this because of knowledge silos and the erosion of team cohesion. But the silos didn't materialize, because getting up to speed on a colleague's work now takes minutes: you ask Claude how something works, where it stands, and what's left, and you're contributing shortly after. His reasoning is that context switching has moved up a level. You no longer need to absorb the specific code as much as the problem and the rough shape of the solution. He's careful to say there may be long-term costs they haven't seen yet, but notes that this is a weaker worry than the confident "we'll regret this" he would have offered a year ago.
The third change is in his advice to the engineers he manages. He used to recommend the T-shaped path: a broad base, then deep expertise in one area such as performance, accessibility, or a framework like Rails. Deep knowledge of a specific technology now matters less, because an AI collaborator can answer detailed questions about any framework's source in seconds. He still wants engineers to build strong fundamentals, especially architecture, abstraction design, recognizing when an abstraction has gone weak, and naming missing concepts, since that's where he thinks human judgment adds the most. But he also says technical skill alone is no longer enough, and that engineers need to learn design and product thinking. The piece is less a set of firm conclusions than an open question about how much of his accumulated experience still applies.