November 21, 2026
Heard in AI

Anthropic's platform team on how it keeps Claude slop in check

Anthropic's Claude Platform team said authors own the quality of anything Claude helps produce, and readers should return real slop but not nitpick Claude-isms. They said the team still debates the balance.

A briefing reports one development at a point in time. We may correct or clarify it later; a new development gets a new briefing. How our formats work

Based on Latent Space, episode published October 27, 2026

People at Anthropic can use Claude about as much as they like. On the Latent Space podcast, members of the team behind the Claude Platform were asked how they keep that from burying colleagues in "slop": low-quality, machine-generated work that someone passes along without checking. Their answer set rules for both sides of a handoff. The person who produces something answers for its quality. The person who reads it should send back real slop, but shouldn't dismiss careful work just because Claude visibly helped write it.

The episode was published on October 27, 2026. Its guests were Angela Jiang and Katelyn Lesse, who work on the Claude Platform, the APIs and agent tools developers use to build on Anthropic's models.

Documents: the author owns the quality

The answer began by calling slop a live debate inside the team, and one Anthropic shares with plenty of other companies. Using Claude to help write a document or code is expected. The trouble starts when someone has Claude generate a piece of work and never reads it closely. The effort of making sense of it then falls on whoever has to use it, which the answer called "insane and totally unfair," and the team pushes back on that habit. Jiang put the rule in one line: "You're still responsible for the quality."

Readers can cause problems too. Claude's writing has recognizable habits, such as an em dash dropped into the middle of a sentence, and readers react as soon as they spot one. They may be put off even when the author spent real time on the work and used Claude only for help. The discussion called that reaction unproductive as well.

So the team's standard has two halves. Authors should use Claude, since it would be silly to ignore what the answer called a "hilariously powerful agent," but the result has to clear a quality bar. Readers who receive actual slop should send it back rather than waste their time on it, and they shouldn't fuss over minor Claude-isms. The answer also acknowledged that the balance isn't settled: the team keeps going back and forth on it, like everyone else.

The conversation then turned to a tease: Anthropic's staff have free tokens (the units of text by which AI usage is metered and billed), more than anyone else, so they probably see far more of this than other teams do. The reply granted the point but said that turning out unreviewed slop is "heavily discouraged" there anyway.

Code: humans design, Claude implements

For engineering work the principle is the same, but the answer described the team as being in a good place. People focus on whether a system's design is sound, and Claude usually writes the implementation. Much of that code needs only "a quick stamp" rather than a line-by-line review, because the overall plan matters more than a perfect implementation and a lot of automated verification (tests and other checks) stands between the code and what ships.

Code slop comes up less often now, the answer said, partly because Claude has improved and partly because the right checks are in place. The exchange closed on a three-word summary, "Trust your tests," met with a yes.

Product: more features is the wrong goal

The question then moved to product work, where agents make it easy to ship feature after feature. The conversation turned to a tweet arguing that the right use of more capable models is not to push ten times more features into production. The time saved, it argued, should go into understanding users, running experiments and thinking carefully about the first prompt before work begins.

The reply came as what its speaker described as a recurring rant. The basics of planning and shipping software haven't changed just because agents speed up parts of the process. Yet some people now treat whole steps as something they no longer need to think about because an agent is handling them, an "almost laziness," when those steps are still the foundation of the work.

The discussion also admitted the pull in the other direction. To a founder, experimenting now costs so little that it feels as though one should simply do more of it. The tweet's advice was to aim the tools at a goal chosen from the start. In the conversation, that advice was called convincing but hard to follow when "you have this toy that does everything" and the temptation is to throw it at the wall.

Example: the claude.ai speed sprint

As an example of high output that still kept quality high, the conversation cited Anthropic's two-week push to make claude.ai, its chat app, faster. Twelve of 13 targets were hit within three days, and the participants noted that this was refactoring (reworking existing code) rather than new product.

Anthropic's engineering report on the sprint, dated September 23, 2026, confirms that 12 of the 13 initial targets were reached by day three of the August sprint. Comparing real-user timings at the 75th percentile on August 13 and August 27, the time until a fresh page could accept typing fell from 3.1 seconds to 0.55 seconds, and the team reports about a threefold improvement overall. Tests, feature flags, human approvals and gradual rollouts guarded more than 3,000 changes, and the authors report no customer-facing incident or rollback. They list performance at the 95th percentile and long conversations as areas still needing work.

The report also shows where people stayed in charge. Humans decided the scope, the order of the work and any tradeoffs users would notice. One proposed 900-line change promised a two-millisecond gain; it was rejected because the speedup didn't justify the extra code to maintain.

Share this article

Go to the original

Sources & further reading

  1. 01

Connected ideas and articles

From the conversation

Podcast episodes

Latent Space

How Anthropic ships the #1 LLM in the World safely: Separating the Brain from the Hands

Episode published This article draws on 1:20:53–1:26:56 (approximate times)

Article history

Updates to this article

Tags

Modal's CEO estimates agents help write 95% of its merged code

Erik Bernhardsson, Modal's CEO, estimates that agents write about 95% of the code merged there "to some extent." He says that, contrary to his expectations, very senior engineers rather than junior ones have embraced agents most, that humans still need to supervise them closely, and that demand for engineers will keep rising.

5 min read

How Tenex keeps its coding agents' context in the repository

Tenex, which builds AI systems for large companies, says it stores plans, agent decision logs and project context beside its code so agents can read them. In the episode, Tenex said Opus 5.5 is the main coding model for most of its team, and that pairing it with Codex as a reviewer still gives the best results.

7 min read