Mike Krieger changed from breaking engineering work into tasks to describing an end state and letting Claude work through the implementation.
2
Anthropic Labs uses two-week persevere-or-pivot reviews, with temporary teams forming around bets and projects being shut down regularly by design.
3
The hardest limit on AI-generated code is often whether a person can understand the change, so reviews need explanations of intent, trade-offs, and production behavior.
Summary
Mike Krieger describes how his work changed after moving from Anthropic's chief product officer role into an individual contributor position at Labs. He now gives Claude a goal and lets it work through the details, rather than specifying every engineering step. He argues that people should ask more of AI systems, especially as they gain access to tools and environments. He gives a concrete example: over one weekend, Claude ported a few hundred thousand lines of Python to TypeScript, checked its work, and produced a deployable result. Krieger also discusses lessons from scaling Instagram, including measuring possible failure signals early and building runtime controls and feature flags. At Anthropic Labs, projects are temporary bets reviewed every two weeks, so teams form around projects without permanently changing the org chart. He sees code review becoming a problem of comprehension, with artifacts explaining intent and trade-offs alongside generated code. He ends with advice about taking time off and naming difficult emotions openly.
Krieger now gives Claude goals instead of engineering task lists
Mike Krieger says he spent his first two years at Anthropic as chief product officer, then moved into an individual contributor role because watching other people build with models gave him increasing FOMO. His use of Claude changed at the same time. He used to write a strategy document and ask Claude to critique it, which was useful but different from building directly. With internal versions of what became Mythos and Fable, he shifted from breaking an idea into the steps he would normally take to describing the desired end state. Claude works through the problem, surfaces trade-offs, and returns with an explanation. Krieger sometimes asks it to explain those choices more simply because it can be smarter than he is about the implementation.
People should ask AI systems for larger outcomes as their access improves
Krieger says product teams need to teach people to be more unreasonable in how they use AI. A non-technical colleague recently asked him to change an internal product, and he realized the colleague could simply ask Claude to do it. Earlier AI products put models in a box by limiting their access to tools and environments. A model that can write code but cannot run it makes ambitious requests feel pointless. Krieger points to Claude's ability to use a virtual machine and write Bash as a way to recover when a built-in PDF parser fails, since it can write a script and try another route. Better tool access makes it practical to describe a result rather than negotiate around the model's restrictions.
Claude completed a large Python-to-TypeScript port over one weekend
Krieger's most unreasonable project involved a Labs codebase written in Python, a language he feels attached to because all of Instagram was written in it. He wanted a better deployment path with Bun, so he asked Claude Code to port the whole project to TypeScript. The codebase had a couple hundred thousand lines, making the request sound foolish by older engineering standards. He created a workflow in which Claude ported the code, verified it, double-checked it, and repeatedly worked over its own output. When Krieger returned on Monday, he had a completed, deployable version. He says production data and segmented tests can also help with larger conversions, while the practical challenge is finding a boundary where the work can start incrementally.
Scaling systems need measurements and runtime controls before an outage
Krieger recalls advice from Instagram's early scaling problems, when the service melted during its launch week. The first lesson was to measure anything that might later matter. During an outage, a team needs to know whether a number is normal or high; adding the metric after the incident is too late. The second lesson was to build thoughtful knobs and feature flags. Early Instagram had simple ramp-outs and rollouts, plus dynamic configuration that let the team change runtime settings within seconds to handle load. Krieger sees the same need in AI systems, where teams make many changing trade-offs. Runtime configuration lets them adjust those choices without waiting for a full deployment.
Anthropic teams delegate through shared, proactive model context
Krieger says most internal model use is more like delegation through tagging than interactive work in a CLI. He compares the experience to Midjourney on Discord, where seeing other people's prompts and workflows can expand what users think to ask for. He saw someone assign Claude responsibility for part of a codebase, monitor a feedback channel, take on tasks proactively, fix issues, and respond when an API changed. That made him realize he had been underusing Claude as a glorified coding tool in Slack. The more advanced model is a teammate that holds context and memory and acts proactively. Krieger describes this as a multiplayer, asynchronous way of operating.
Code review is limited by human comprehension, so intent must travel with code
Krieger says Anthropic is still bottlenecked by reviews, especially changes that touch architecture. The deeper problem is whether a person can fully conceptualize what a large change does. A 2,000-line pull request can look like code without giving the reviewer a way to understand the purpose. Claude Code artifacts address this by including the intention of the change, an explanation, and the trade-offs made. Krieger does not read every line of every pull request. He asks Claude about the questions he would have and has it investigate, while keeping the review human-driven for important changes. Cosmetic visual changes can be fixed forward when needed. He expects communication about intent and trade-offs, followed by production measurement, to matter more over time.
Labs keeps projects temporary and org structure stable
Anthropic Labs reviews every project on a two-week persevere-or-pivot cycle. Projects are regularly pivoted or shut down, and Krieger says that is the purpose of the group rather than evidence that the process failed. The team aims to prototype quickly, ship internally, try early access, and wind down work that does not succeed. Since projects change every two weeks, the org chart is not built around them. A bet draws people from product and engineering, with a bet lead who usually manages none of those people. Engineering managers focus on coaching, development, and assigning people to work they care about. When Claude Design gained traction and shipped a second release, it moved from an ad hoc group to a more permanent team.
AI startups can still win through detailed knowledge of a user group
Krieger rejects the idea that startups should give up because Anthropic might release a product that enters their market. He joined Anthropic partly because better models could make experimentation and iteration much faster, not because labs would solve a startup's product taste or ideas. He compares this with Google launching a photos product that would naturally reflect Google's strengths and existing integrations. Labs can provide a platform, while a smaller company can focus closely on a particular industry or group of people. Krieger says the hard parts remain understanding users, reaching them, listening, and iterating quickly. Writing code may affect the timeline, but he does not see it as the thing that determines whether a startup works.
"I think as an industry or even as a product team, we have to teach people to be more unreasonable in their usage."03:02
Who should watch
You are deciding whether to give an AI coding system a small task or an outcome, and want a concrete example of a larger delegation workflow.
Your team is building experimental AI products and needs a way to review projects without reorganizing every time a bet changes.
Large generated pull requests are becoming difficult to understand, and you want ideas for reviewing intent and trade-offs alongside code.
