Silicon Valley GirlClaude Engineering Lead: "Stop Using Claude Like It’s 2024"
At a glance
WHAT IT’S REALLY ABOUT
Use Claude ambitiously: redesign workflows, validate outputs, avoid role-play agents
- Felix Rieseberg argues that many people underuse AI by treating it as a 2024-style helper for small tasks instead of pushing it toward ambitious outcomes and new capabilities.
- He describes his current work as rapid prototyping with Claude—generating many variations quickly—then relying on human attention, taste, and user “click” moments to decide what’s worth pursuing.
- He explains that for creative and knowledge work, “good” is hard to define, so better evaluation often comes from mirroring expert processes and checking methods rather than demanding a perfect final output.
- He shares practical reliability techniques—adversarial reviews and separate “judge” sessions—because a single session is often poor at evaluating its own work.
- He contends that the biggest remaining opportunity is building empathetic, high-value products (especially replacing clunky legacy software), since execution is now cheap but understanding real needs is still hard.
IDEAS WORTH REMEMBERING
5 ideasAim beyond small tasks; use AI for higher-level outcomes and new capabilities.
He argues that simple uses like summarizing emails don’t tap into “new intelligence,” and that the bigger gains come from reframing work around outcomes and letting the model help reach them in ways you might not have been able to do manually.
The real leverage is workflow redesign, not isolated prompts.
He says the main change is not just what the tool can do, but how you redesign the process around it—starting from defining what “good” looks like and building evaluation into the workflow.
Evaluate AI outputs by process quality, not just result quality.
He suggests judging AI work by the quality of its method (e.g., reputable sources, correct steps) rather than only whether the final answer is correct, using the weather example to show why a lucky guess shouldn’t “pass.”
Use adversarial reviews to catch hallucinations and overconfidence.
To protect against confident mistakes, he recommends running adversarial reviews: generate the output with one session, then spin up multiple other sessions whose job is to find flaws and prove it wrong.
Separate creation, critique, and judging into different sessions.
When adversarial feedback is noisy or conflicting, he describes using a separate “judge” session to validate whether criticisms are legitimate, and optionally routing valid feedback back to the original session.
WORDS WORTH SAVING
5 quotesIf all we're doing is, like, summarize this email for me, or like, make this email sound less angry, you're not really, like, benefiting from the new intelligence.
— Felix Rieseberg
I'm just a conveyor belt for Claude, right? Like, I go look at our crashes, pick one up, move it over, drop it off in front of Claude, and I'm like, "Please fix this."
— Felix Rieseberg
Adversarial reviews basically boil down to the idea of you've given, given one Claude a task- Spin off another 10 that try to, like, break this down. Their whole task is to attack this and prove to you why it's wrong.
— Felix Rieseberg
I rarely want the current session to judge itself. That I think is still really iffy.
— Felix Rieseberg
I think the secret to building a good product has actually not changed, and I believe it's empathy.
— Felix Rieseberg
High quality AI-generated summary created from speaker-labeled transcript.
