Anthropic Head of Engineering on Stopping AI Slop

Key Takeaways

• Katelyn Lesse requires Anthropic engineers to read and edit AI outputs before sharing them, preventing reviewers from wasting time parsing unvetted text.
• Anthropic shifted code review from line-by-line syntax scrutiny to verifying high-level system plans, leaving implementation details to Claude.
• The team used Claude to rewrite core Python services into Rust, proving that AI-assisted rewrites succeed only when humans own the architectural boundaries.
• Latent Space host swyx and Lesse argue that faster coding should buy time for customer discovery and testing, not a ten-fold increase in surface features.

The Cost of Unread Code

When engineers use Claude to draft an engineering spec or write a pull request, the easiest failure mode is clicking send without reading the text. Lesse sees this pattern as a transfer of work from writer to reader. If an engineer generates a full design document without proofreading it, teammates spend their afternoon hunting for subtle hallucinations and awkward phrasing.

“If you use Claude to totally generate a thing and you didn't spend any time actually reading it yourself and making sure that it was of high quality, then you're putting a whole extra burden on the consumer of that document or of that code to spend time trying to parse it,”▶ 1:21:31 Lesse explains. “You as the human are still responsible for the quality.”▶ 1:22:00

At Anthropic, teammates spot unreviewed outputs quickly. Stylistic quirks give them away immediately, including telltale punctuation patterns like unexpected dashes in a document. Once readers spot those patterns, trust in the technical proposal drops. Lesse points out that the goal is not to hide the use of Claude, but to ensure that the human whose name is on the pull request actually evaluated the logic.

Review the Plan, Not the Syntax

Software teams traditionally spent hours checking variable names, edge cases, and line-level syntax during code reviews. With models handling routine implementation, that time allocation makes little sense. Anthropic restructured its engineering expectations around architectural plans rather than individual lines of code.

“Humans play a large role in making sure designs of systems are good, and implementation tends to be Claude,”▶ 1:23:56 Lesse notes. “For a lot of the code that gets written, all you really need is a quick stamp and you can keep going. You don't necessarily have someone deeply reviewing every single line of the code because you care more about the overall plan.”▶ 1:24:05

This shift proved itself when Anthropic ported major internal components from Python to Rust using Claude. The engineering team did not audit every auto-generated Rust function from scratch. Instead, senior engineers designed the safety boundaries, concurrency models, and memory layouts. Claude produced the typed code to match those specifications, and automated test suites verified the behavior.

Accelerated generation creates a new temptation: building too much junk. As swyx pointed out during the conversation, “The right way to use model capabilities is not to ship 10x more features into prod. Spend more time understanding your users, trying experiments, building whatnot. You could very easily just keep shipping random slop, but you should actually think about it.”▶ 1:24:53

Lesse agrees that developer laziness is the main risk for engineering orgs. When agents write functions in seconds, engineers stop interrogating their assumptions. Anthropic counters this trap by demanding that engineers spend their saved hours on rigorous user research, failure-mode analysis, and clean system design.

What to Do With This

Audit the pull requests your team submitted this week. If any pull request contains raw model outputs that the author cannot explain line by line in a five-minute design review, reject the diff. Mandate that every AI-generated document includes a two-sentence summary written entirely by hand before teammates open it.