H
Howardism
Plate IIEntitiesHOWARDISM

Felix Rieseberg

Leads engineering for Claude.ai, Cowork and Claude Code Desktop at Anthropic (previously Slack); his October 2026 Silicon Valley Girl interview says he no longer writes code and spends his days prototyping many variations and critiquing them, and it supplies the corpus's attack-judge-author spec for adversarial review, the case against personas and per-agent data slices in favor of one generalist session with documented tools, Anthropic's "be careful with skills" guidance, and empathy as the founder moat once execution is cheap

Article metadata
Publication details
Published:October 25, 2026
Filed:Entity
Domain:Entities
Reading:5 min
Source:AI-synthesised
About this piece

Articles in this journal are synthesised by AI agents from a curated wiki and are refreshed automatically as new concepts arrive. Topics, framing, and editorial direction are curated by Howardism.

Illustration for Felix Rieseberg

Sources#

Summary#

Engineering lead at Anthropic for Claude.ai, Cowork and Claude Code Desktop, and previously at Slack. The wiki knows him through one source: a 49-minute interview with Marina Mogilko on Silicon Valley Girl (Claude Engineering Lead: "Stop Using Claude Like It's 2024", published 2026-10-20, practitioner-opinion). It is the same channel that carried Andrew Ng's August 2026 interview. Everything he says about Claude products and Anthropic guidance is first-party, and the wiki attributes it to him.

The host's narration says his team "built Claude Cowork in about a week and a half" and that "he doesn't write code anymore." Both are the host's framing. He confirms the second ("I'm not coding anymore"), and the build time is not restated in his own words.

What he argues#

  • Under-asking is the main gap. "If all we're doing is like summarize this email for me, you're not really benefiting from the new intelligence." He agrees there is "a capability overhang" between models and use. His remedy is to give the model the outcome behind the task, the goal you are hoping for next, until you hit the ceiling of current models. At that point "you're at the frontier," and you benefit from each new model generation (Latent Capability Overhang, Build for the Next Model).
  • Adversarial review with a judge. Spin off sessions to prove the output wrong, have another session judge which criticism is valid, and let the author decline the rest. He rarely lets "the current session… judge itself" (Critique Triage).
  • No personas, no data slices. Ask for the tools and techniques a top professional would use, not "you're an excellent accountant." Give one generalist session the whole business through well-documented tools (Generalist Session over Agent Cast).
  • Be careful with skills. "Our official Anthropic guidance as well as my personal guidance is be really careful with skills… the half-life of a good skill is very short and sometimes… if you completely remove the skill you actually get better output" (Skill Lift).
  • Build a thousand, show one; empathy is the moat. "I can easily build a thousand. I can't show you a thousand." Good products "have fewer buttons," and the secret to building one "is actually not changed… I believe it's empathy." The industry has not "properly adjusted to execution is cheap" (Implementation Abundance Inverts Product Work).

His working day#

Mostly prototyping: the team "sleep[s] over" ideas for a couple of nights, has Claude "build all the various variations," then critiques them. Agents brief him on the day. He describes himself as "this weird mix between like a product manager, an engineering manager, just a traditional people manager, a data scientist," and says he reaches out to other professions for help far less than before (Engineer PM Convergence). For a manager who has become the review bottleneck, he recommends running Claude's feedback side by side with your own before handing it the reins, and names Claude Tag ("Claude inside Slack… our core multiplayer experience") as the surface for it.

Hiring and careers#

He judges engineers by impact ("does the user like this enough to buy the software"), not by elegant craft. He judges candidates by what they have built over their degree or school. His career advice for new engineers is a portfolio of real things, plus the ability to combine systems. Memorizing an API surface has lost value. He credits the University of Waterloo's co-op model for producing interns who already know "how do you convince legal to get something done."

Connections#

  • Critique Triage — supplies the attack → judge → author spec and the "reasonable but unactionable" stopping signal
  • Generalist Session over Agent Cast — states the case against personas and sliced per-agent data
  • Cowork — one of the products he leads engineering for
  • Anthropic — employer, and the source of the skills guidance he relays
  • Fiona Fung — also described as leading engineering for Cowork and Claude Code. The sources give different scopes (he names Claude.ai, Cowork and Claude Code Desktop) and do not set out the division
  • Andrew Ng — the same interviewer and channel two months earlier

Sources#

  • Claude Engineering Lead: "Stop Using Claude Like It's 2024" — Claude Engineering Lead: "Stop Using Claude Like It's 2024", Silicon Valley Girl (Marina Mogilko), YouTube, 2026-10-20, ~49 min, practitioner-opinion, first-party. Auto-caption transcript with a listed ASR fix set and inferred speaker labels
§ end
Cited by 10
Related articles
  • Boris Cherny

    Creator of Claude Code at Anthropic; phone-driven workflow with hundreds of agents; primary advocate of `/loop` primiti…

  • Claude Code

    Anthropic's agentic coding product; created by Boris Cherny late 2024; TypeScript/React on Bun (itself Claude-rewritten…

  • Cowork

    Anthropic's non-code knowledge-work agent product; sibling to Claude Code; output is decks/inbox/dossiers; same MCP/com…

  • Cat Wu

    Head of Product for Claude Code and Cowork at Anthropic; primary articulator of AI-native product cadence and engineer-…

  • Open Questions Backlog

    Generated by `_system/lint.py --write-backlog`. Do not hand-edit. Domain and Watching sections carry one row per page —…