November 21, 2026
Heard in AI

Fitting Claude to every user is "super hard," says Anthropic's Jiang

Anthropic's Angela Jiang says early-access API customers catch problems internal testing misses, but she calls tailoring Claude to users with conflicting preferences "super hard" and says its gentle defaults work only "okay-ish for now."

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

Some people who use Claude want it to stop being lazy and finish the job. Others complain that while they were still sketching out a plan, it went ahead, wrote a pile of code or took actions they weren't ready for. On the Latent Space podcast, Angela Jiang, who leads the product side of Anthropic's team, conceded that the company has no good answer yet for fitting Claude to each user's preferences. Asked how Anthropic squares one default behavior for everyone with what individual users want, she said: "We don't do a great job of it. Seriously, it's really hard. It's super hard."

Jiang and her Anthropic colleague Katelyn Lesse appeared on an episode published October 27, 2026 about the engineering behind the Claude Platform, the tools Anthropic offers developers who build on its models. The host asked what goes into a model launch while that work was still fresh in their memory. Anthropic had announced Claude Opus 5.5 on September 22, 2026.

From a trained model to early access

Jiang said launches have become "a bit of a machine," apart from the safety work, which she called the more complicated part. Anthropic makes the model, tests its safety, evaluates it against where the company thinks it needs to be and makes tradeoffs. Then it cuts several versions, or variants, and evaluates them. Anthropic staff also use the variants heavily for their own work. Jiang said that internal use catches problems early and shows which variants are better than others.

Once the team feels confident about a set of variants, it gives one to customers. Sometimes that happens through Anthropic's own products, such as the coding agent Claude Code or Cowork. Other times it goes to API customers, the companies that call Claude from their own software. Jiang described this as effectively an early-access program and named the AI coding companies Cognition and Cursor among the participants. Those customers have to adapt their harness, the software around the model that gives it instructions and tools and carries out its actions.

Anthropic then asks them questions: Does the model perform well in your evals? Do you like it? What don't you like? Evals are automated tests that score a model on tasks a customer cares about. Some of the feedback is "vibey," Jiang said, and Anthropic wants that too. When everyone feels good about all of it, the model launches.

Jiang said API customers are especially useful because they put Claude into a different harness. Anthropic wants Claude to be very general, so it matters how well the model works inside other companies' software, and that feedback differs from what Anthropic's internal use turns up. Praise mostly confirms what the team already suspected. "Honestly, the most useful feedback is negative," she said. When a customer reports that the model didn't do well at this or that, the team's next question is how to fix it.

Feedback on a tight clock

For all that process, the conversation acknowledged that model feedback is still loosely organized. It often runs through a Slack channel shared by the two companies: one side reports that something seems weird, and the other asks for an example it can fix. The explanation given was speed. Anthropic wants the gap between cutting a test version and opening the model to everyone to be as short as possible, so much of the exchange happens on the fly. That is why it chooses early-access partners with strong eval suites, who can run all their tests quickly and come back with actionable feedback along with their impressions. The discussion also noted that surprisingly few agentic products, meaning software in which the model takes multistep actions, have evals good enough to help test a new model.

Too lazy, so back to training

The conversation then turned to laziness versus proactivity, a spectrum on which different users want Claude in different places. Because people want different things, getting the balance exactly right was described as impossible. In one case, an early version of a model showed too much laziness, and Anthropic sent it back for more training to remove the problem. The discussion presented this as the kind of issue that internal testing within Anthropic's own products and API might have missed, which is why the company puts early versions in front of API customers and asks them what needs to change.

Alignment versus personalization

The question put to Jiang framed this as alignment versus personalization, probably one of the hardest parts of managing a model as a product. Anthropic wants a sensible default personality and default behavior for everyone. It also wants Claude to respond quickly when a user's CLAUDE.md file says something like: please be more proactive than the default. CLAUDE.md is a file of project instructions that Claude Code reads. According to Anthropic's documentation, people use it to record things like test commands, architecture and coding conventions. Its contents go into the model's context as guidance and are not mechanically enforced. AGENTS.md is a similar open convention for giving instructions to coding agents.

The example raised was FrontierCode, a coding benchmark from Cognition; the questioner said they had been involved with it. Cognition says the benchmark grades changes drawn from real open-source pull requests against criteria set by reviewers, and a run that fails a blocking requirement scores zero. FrontierCode deliberately penalizes too much activity, such as touching too many files. But a user could simply tell the model in CLAUDE.md not to do that, and it wouldn't. If the benchmark has no room for that kind of personal instruction, the question went, is it penalizing Opus 5.5 unfairly?

Anthropic's constitution came up as one answer. Anthropic describes the constitution as a training document that explains how Claude should behave and why, with broad safety, ethics, compliance with Anthropic's guidelines and genuine helpfulness as its priorities. Jiang said "the constitution's helpful," but that instructions and preferences vary so widely that adapting to each one is very hard.

Her team's main tool is parameters, settings that let Anthropic adjust Claude for each customer base. When Anthropic doesn't control the end users, as with API customers, it gives recommendations and leaves the adjusting to the customer. The settings cover things like how much effort Claude should put in and what size of model to use. Jiang said this does not reach the level of personalization that a CLAUDE.md or AGENTS.md file offers. Beyond the parameters, the team weighs the conflicting tastes in all the feedback it receives and makes judgment calls. Jiang said that approach won't scale as Anthropic adds customers and users. She predicted Anthropic will need ways for Claude to take in a user's personalization, perhaps through a signal the user provides or something they could tune on.

Effort set differently for different users

Effort is one of those settings. Anthropic's effort documentation describes it as a control over how many tokens, the units of text a model reads and writes, Claude spends on thinking, explanations and tool calls. Lower settings can mean fewer or terser tool calls, and higher settings put more resources into hard work. The documentation lists medium as the default for Opus 5.5. In the episode, the conversation noted that the default effort had changed to medium, and Jiang confirmed that this was a Claude Code setting.

Jiang called effort a perfect example of a parameter Anthropic uses to "customize" Claude for different groups, and said "personalize" might be too strong a word. The default changes depending on whether someone is a free or prosumer customer, uses Claude Code or Cowork, or is an enterprise or API customer. Users can always change it. She said Anthropic could take a heavier hand but doesn't want people to feel they can't control their own setup. Some users want the high end of Claude's intelligence; others want to try the low end when it is faster for their task. Jiang called that a hard product tradeoff. Anthropic leans toward "gentle defaults" with options to adjust, which she said works "okay-ish for now."

A growing tangle of settings

The conversation acknowledged that Anthropic isn't very happy with any of this yet. Each launch raises the same question: when Claude gets good at something new, should the API expose it as a parameter that customers control? Over time, the answer has produced an odd mix of effort levels, budgets and other settings, and it is hard for anyone to know which combination suits a given task. The open questions named were which parameters to retire, which to give better defaults, and whether a setting should use a 1-to-100 scale or low, medium and high. Claude has been trained to understand a whole range of options, but the discussion concluded that exposing every possible combination would not get customers very far.

The answer the episode offered was Claude Managed Agents, which gives programmatic access to Claude through an API but comes with a ready-made harness. The effort documentation notes that Managed Agents can set effort in an agent's definition or override it for a single session. With every model launch, Anthropic does the work of finding the new parameters and breaking changes and figuring out how to get the best from Claude inside that harness, so the people using it don't have to think about it.

Share this article

Go to the original

Sources & further reading

  1. 01
  2. 02
  3. 03
  4. 04
  5. 05
  6. 06

Connected ideas and articles

Background

Stories behind this article

  • Anthropic releases Claude Opus 5.5 The recent Opus 5.5 launch sets up the article's account of how Anthropic launches models. Opus 5.5 also comes up in the FrontierCode question and the effort defaults. The article is not about the release's pricing, capabilities or adoption.

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 30:07–41:35 (approximate times)

Article history

Updates to this article

Tags

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