Claude Engineering Lead Felix: Stop Using AI Like It's 2024 — Ten Methods That Actually Work

Felix Rieseberg, engineering lead for Claude at Anthropic, recently shared how he works now. His team built Claude Cowork in about a week and a half, and he himself barely writes code anymore — he calls himself "just a conveyor belt for Claude." His core test: if you're only using AI to summarize email and polish prose, you're not benefiting from the new intelligence. What pays off is redesigning your workflow so AI touches raw data and real tools. Ten takeaways worth keeping.

1. Stop being the conveyor belt — grant full data access

A year and a half ago, when Felix hit a crash, he'd carry it to Claude one by one: "fix this." He quickly realized he was just a conveyor belt. The right move is giving Claude access to all the crash data and letting it decide what needs fixing. The setup costs more up front, but once built, "it opens a lot of doors."

2. Know the five parts of an agent

The model (perception, reasoning, decisions), the harness (coordinating memory and tool activation), the sandbox (a safe runtime), knowledge (company data turned into a knowledge base), and tools (the means to actually do work). Knowing these five tells you where to spend effort and budget.

3. Don't role-play — demand professional process

"Act like a great engineer" is a 2024 relic; the model just performs vocabulary. His preferred prompt: "Tell me which tools the top people in this field use, then use those same tools." The point isn't the costume, it's the professional toolkit and method.

4. Adversarial review: one does the work, ten attack it

To keep AI from confidently handing you wrong answers: give one Claude the task, then spin up ten more whose only job is to attack the output and prove it wrong. Finding flaws is easier than creating, so review works even a tier down in capability. He runs five threads in Claude Projects, then has another Claude act as judge on which criticisms are valid; valid feedback flows back to the original session. Key caveat: a session is bad at judging its own work — seek hard external signals.

5. Give agents the full business view, not data slices

Per-agent datasets and "personas" are over-rotation. Models are strong generalists and tool calling has matured. The right pattern: when data is large, give tools and document what data lives behind each one — the agent fetches what it needs instead of you pre-deciding its slice.

6. "Model experience": pretend you're the model

When designing a workflow, ask: if I were Claude, what would I want to complete this task? The answer is usually raw data, a computer, freedom to install software, and room to decide — not pre-chewed conclusions.

7. Skills have short half-lives

He still uses skills, but carefully: many skills encode advice that only fit the previous model generation, and AI time compresses fast. "A good skill has a very short half-life." Sometimes deleting the skill and prompting directly produces better output — provided you have evals and know what good looks like.

8. AI products are in the "pre-iPhone era"

Today's AI products are like the motley phones before the smartphone. Don't mistake the current state for the end state. His trust-building method: run side by side — you read, write your feedback, do a few rounds, then compare with Claude's. The differences mark exactly where your workflow needs sharpening.

9. Execution got cheap; empathy and judgment got expensive

In 2015 a study-abroad booking platform meant hiring engineers and burning hundreds of thousands of dollars. Now he could rebuild it in two days. As execution cost collapses, the secret of great products is unchanged: empathy for users and scenarios, and judgment about what's worth building — "the MBA's revenge," he calls it. Personally: keep handing tasks to the model until you hit the ceiling of current intelligence. Hitting the ceiling means you're at the frontier and will benefit from the next model generation.

10. Hire portfolios, not degrees; politeness to models is for your own soul

Always pick the person who has already proven they can build something remarkable over the one with a good degree and no artifacts — and today's junior engineers are often more AI-native than veterans. As for saying "please" and "thank you" to models: he does, because "as a human, keeping a little respect toward machines is better for me, for my soul."

Closing

The most memorable thing here isn't any single prompt but a self-test: are you "using AI," or "pushing AI to the edge of its capability"? The former treats Claude as a smarter chat box; the latter treats it as an engineering system to be designed, constrained, and verified. In 2026, the right posture is making AI run into your current intelligence ceiling.

Source: Felix Rieseberg interview.