Claude engineering lead Felix Rieseberg on prototyping, adversarial review and what makes a good product

The conversation opens with Felix Rieseberg, whose team built Claude Co-work in about a week and a half. He now leads engineering for Claude at Anthropic and, by his own account, no longer writes code.

He describes his job in terms of directing the model rather than typing software. Asked about the work, he says: "I'm just a conveyor belt for Claude." In his account, Claude needs access to data such as crash reports, and then works out for itself what needs fixing and what does not.

Asked whether there is anything he would not trust an AI agent with, he says he does not yet trust a session to judge its own work and calls that "still really iffy."

Rieseberg says most of his day goes to prototyping. He and his colleagues generate ideas, decide which are worth pursuing, and then ask Claude to build the variations and critique them.

Asked how to decide whether an idea is good, he says he does not have all the answers. He argues that the calculus has changed. Engineering teams used to weigh maintenance cost, ongoing cost, time investment and opportunity cost before building anything. Now a prototype can be reached much faster.

He says human attention is still limited. He could build a thousand things but can show someone perhaps one a day, so he believes good products have fewer buttons, and a product works when people quickly do things they could not do before.

Asked how to define good text for knowledge work, where bugs do not provide a clear test as they do in coding, Rieseberg uses a weather-reporting example. If a model simply guesses the temperature and happens to be right, he says that should not be a passing grade.

In his view, a model that is off by half a degree but found the right location and reputable data sources, with a reputable process, should be judged better than a lucky guess. He notes that writing and other outputs where taste matters are harder to judge.

The host suggests that AI might mimic the steps of the best people in a field, but that taste would still be missing. Rieseberg agrees that such work involves a lot of effort, but says that when he talks to good writers, most of them would disagree that their text is good because of a magic touch.

He says many artists are proud of a process refined over years, involving many manual steps, rather than believing they sit down and produce a great work at once. He points to writers' studios full of whiteboards and information as evidence.

He notes that Claude responds well when told to take the steps a professional in a field would take, rather than to produce a result that merely looks right. He says he sees the most progress in motion video and 3D work, where he believes Opus 5.5 has made a pretty big jump. But he adds that the bigger leap for the public has come from simply telling Claude to use Blender, the same tools professional animators use, after which the results are quite good.

Asked how to protect oneself from AI confidently presenting mistakes, Rieseberg says this still happens to many people and offers what he calls a practical answer: adversarial review. He says it has proven effective in coding, math and other areas where results can be measured numerically.

In the method he describes, one Claude is given a task and about ten others are spun off with the job of breaking it down and proving why it is wrong. He compares this to red teaming or short selling, and when asked whether the same model should be used for that, he says it depends on the task.

He says it is sometimes easier to find flaws than to create something new, just as it is easier to judge whether a new phone is good than to design one from scratch. So, he argues, a step down in intelligence can still produce very good results. Asked how he uses this himself, he says he asks Claude to start five threads in Claude projects, each tasked with proving that a blog post he wrote this morning is wrong, including whether he has his facts or reasoning wrong.

Rieseberg sorts the question into two cases. If you have no idea what good looks like, his answer is more AI, throwing more tokens at the problem. If you do understand what good looks like, that may be enough, and you need not hold opinions about the process itself.

The conversation turns to letting go. The host suggests keeping a close eye on only the most important projects, and Rieseberg agrees that it is hard not to micromanage.

Asked whether an agent needs only the data it uses, given how much data there is, Rieseberg says Claude is discerning about the data it wants. He argues against dividing data into slices for each agent, because people who want to be effective need the full view of the business.

He compares this to an engineering manager who has only code access and is therefore much less useful than one who also has customer data and access to customer calls. To handle the volume, he says, the answer is tools, documented well so the agent knows what data sits behind each tool and does not have to ingest everything.

Rieseberg says his team still uses skills, but Anthropic's official guidance and his personal guidance are to be careful with them. Prompts often carry guidance that was relevant to the previous model generation, and in AI that time is compressed, so the half-life of a good skill is very short.

He says that sometimes removing a skill entirely produces better output. He suggests that this is most useful for people who have an evaluation and a clear sense of good and bad results, and that he is discerning about when he uses skills. He proposes thinking about the model's experience rather than the user's, asking what a model would want in order to complete a task, such as raw data, access to a computer and the ability to install software.

Asked whether he has dedicated agents for research or tax returns, Rieseberg says he has projects and thinks about different work items, but not different roles. He says he never tells Claude "You're an excellent whatever," a style he associates with 2024, when people told Claude it was an excellent engineer and asked it to fix problems.

He says he asks Claude to name the tools someone at the top of a field would use, and then to use them. He argues that a common problem is role play, which models still tend toward if told to act as a very good accountant. What he wants is the techniques, tools and knowledge a professional would use, not the vocabulary or costume.

Rieseberg says there is no agent he trusts 100 percent, so every output goes through his team and eventually through him, which makes him the biggest bottleneck in the company.

The host says the products could handle much of that feedback today and suggests running the agent side by side with one's own reviews. Rieseberg says he recommends this experiment to almost everyone. Comparing the two sets of feedback shows where they differ and helps sharpen the workflow, because even people give different feedback on the same document on different days. The host suggests Claude Tag, Claude inside Slack, for building such an agent for a team.

Asked about his entrepreneurial moat when software is cheap to build, Rieseberg says he thinks it has not changed much. The upfront capital cost was much higher before, and single-person or two-person teams can now build billion-dollar companies.

He says the secret to a good product has not changed, and he believes it is empathy: for the user, the use case, and whether something is valuable to the person who will pay for it. He adds that this has always been a key ingredient in entrepreneurial teams, and that investors often regard the founder as more important than what they bring.

Rieseberg says that when he visits airports and hotel check-in counters, he sees systems that have not moved to graphical operating systems and still show black-and-white screens. He describes the software on these devices as clunky and awful.

He says he thinks this is probably an untapped opportunity for many businesses to clean up. He recalls that years before AI, he joked that if he ever wanted to retire he would build good software for dentist offices, and says it now could be done in a day.

The host says a Nobel laureate in economics told them models are very powerful but the application layer is lacking. Rieseberg agrees there is a capability overhang: someone with an advanced model may only ask it what the capital of France is and miss what is possible.

He cites CEO Dario Amodei's idea of a data center of geniuses and says his job is to work out how to empower users to make use of that. He argues this takes creativity and a rethinking of needs and of what seems possible. He says people have always been held back more by what they think is possible than by what is actually possible, and that the advice to be more ambitious is context dependent but means shedding beliefs about what is possible and rethinking why one does things.

Rieseberg says his recommendation for young people is to build real things and prove they have empathy and the ability to talk to customers, figuring out what is necessary and what is not. He describes this as a slightly new definition of software engineering, which he thinks is good.

He says younger people adapt to this style of thinking faster than some older engineers who enjoyed working on tickets without having to ask why too often. He would no longer recommend memorizing individual API services, which he says has become less valuable. What remains valuable is combining different systems in interesting ways, and AI amplifies that by letting someone point it at three or five systems and have them combined.

Asked about hiring, he says he still looks at resumes, and that relationships matter in his industry because he wants someone who knows how other businesses operate and what is happening in the market. For creative roles he looks at how people work and what they know about the art.

Rieseberg says he would advise his daughter to understand early how businesses operate, how people collaborate and how to find real value. He admires the University of Waterloo's approach, which he sees as an outside perspective. From Germany, where he says computer science is largely informatics with much math and little on building a product, JavaScript was not considered real computer science.

He says Waterloo co-op interns often arrived having worked at several companies, sometimes four internships in and with stints at firms such as Tesla, Google and Apple. Those interns learned how companies operate, the upsides and downsides, how to advocate for an idea, how to get resources and how to convince legal. He says these skills made Waterloo engineers almost immediately effective, placing them in a different category from someone with a computer science PhD who has never worked in a company.

Rieseberg says he told his daughters that anything is possible with hard work and passion, having seen that the people he met were normal but very passionate. The host agrees that AI discussions often focus on ambition, and that people do not try what they do not consider possible.

Rieseberg says that what people believe is possible is a bigger limit than what models can do or what people can do. He says he also believes in the new generation because they will not have experienced the limitations older people carry.

Asked about the next five years, he says forecasting is an occupational hazard and that the biggest thing he has noticed is that everyone is wrong about the intelligence exponential coming from AI. He recalls that when Google's AI team won a silver medal at the math olympiad, he was well aware of scaling laws and would still not have predicted that AI would solve otherwise unsolvable math problems as early as the middle of 2026, when he would have guessed it was a year and a half away.