Anthropic engineering manager Felix Rieseberg wrote on his personal blog that, starting this month, he has stopped using traditional coding interviews when hiring. His reasoning is blunt: the things these questions measure are things Claude already does better than candidates.
Rieseberg has led the engineering teams behind Claude.ai, Cowork and the Claude Code desktop app at Anthropic. He is also a co-maintainer of Electron and previously worked at Notion, Stripe, Slack and Microsoft. He describes the coding interview as a “weak but useful” predictor, a rough signal of who will do well on a team, and says that signal has now stopped working.
“Claude is now better at both”
The line comes from a comparison he draws between a React coding challenge and a quiz on assembly: whether the test covers a front-end framework or low-level details, the model answers better than people do. He also writes that the people with the biggest impact on his team have no connection to how well they score on such tests, and that coding interviews, although used everywhere, have never been popular.
Two new hires, one weekend
The core of the post is advice for less experienced engineers, under the title “Hunger beats tenure.” His example is the Claude Code graphical interface. Engineers had long been used to Claude Code existing as a command-line tool, and nobody had built a GUI. Two recently hired engineers, Anthony Morris and Alice Zhao, put together a working prototype over one weekend, and the first public version shipped about eight days later.
His advice boils down to five points:
- Work out what the company is trying to achieve, ideally understanding the goals your manager's manager cares about, then pick work that moves the business instead of grinding through a backlog;
- Come up with a few bold ideas around those goals;
- Build a prototype over a weekend, skipping the requirements doc, the lobbying for support and the scheduling of designers and product managers;
- Put just one idea in each prototype, so leadership can see the opportunity and also see that you can carry it forward;
- Use the newcomer's perspective: people who haven't heard all the reasons something “can't be done” are often the ones who get it done.
He sums up the shift behind this in one sentence: execution is no longer expensive. The framework knowledge accumulated over many years used to determine how much impact one person could have; as models bring down the cost of writing code, that barrier has loosened. At Notion he was once told offline mode couldn't be done. He later found that the complexity of the data model was real, but that the founders' supposed opposition was just something everyone assumed.
What this means for hiring in China
Rieseberg also admits he hasn't found a better interview format. His current approach is to go through résumés and portfolios looking for small prototypes that show “taste, empathy and principled thinking,” and he says he doesn't expect coding interviews to come back. This is one manager adjusting his own team's process; Anthropic has not announced a company-wide change to its hiring, and it cannot be verified whether other teams will follow.
For engineers in China, this logic is quite far from current practice. Large Chinese tech companies generally keep online coding tests and whiteboard-style algorithm rounds in both campus and experienced hiring, and question banks on practice platforms remain a standard part of job preparation. Once models solve problems better than most candidates, test scores will separate fewer and fewer people, and interviewers will sooner or later need other signals. Whether a candidate uses AI during a remote test is also getting harder to police.
The more realistic change may show up first in portfolios. A working small tool, or a string of commits actually merged into an open-source repository, is harder for AI to do on a candidate's behalf than an algorithm score, and closer to what Rieseberg calls “taste.” There is a catch, though: building a prototype in a weekend presupposes understanding what the business needs, and that is a step where models help little.
Sources: Felix Rieseberg's personal blog, CocoLoop editorial compilation; checked the scope of the interview change, the number of engineers behind the Claude Code GUI prototype, and the time to the first public release.