Inside Claude Code: Boris Cherny and Cat Wu on Latent Space
In a May 2025 Latent Space episode, Boris Cherny and Cat Wu of the Claude Code team explain how it started, why it is built like a Unix utility, how much of it Claude writes, what it cost per user, and how they run it headless.
Updated November 20, 2026: Expanded with the episode's main ideas, the practical tips on headless runs, and what has changed since.
Hosts Alessio and Swyx sit down with Boris Cherny, lead engineer on Claude Code, and Cat Wu, its product manager, in a Latent Space episode published on May 7, 2025. The full transcript's timestamps run to 1:13:46. It is the team's own first-hand account of why Claude Code looks the way it does. Some details are now dated: the cost figure, the pricing position and the models discussed all describe May 2025, and we flag them below.
How Claude Code started
Boris describes it as a side experiment with no master plan. He was trying the public API in different places, including a Claude in the terminal that commented on the music he was playing. Once he gave it terminal access and the ability to code, he used it every day. The core team followed, then every engineer and researcher at Anthropic. He recalls the internal daily-users chart going close to vertical for days before the team decided to release it.
Do the simple thing first
The recurring theme is an Anthropic product principle: do the simple thing first. Two examples stand out:
- Memory is a markdown file. After looking at memory architectures and external products, the team shipped CLAUDE.md, a file read into context automatically. It can sit in the project root, in child folders or in your home directory.
- Compaction is a summary. They tried rewriting old tool calls and truncating old messages. In the end they ask Claude to summarise the earlier conversation and carry on from that.
Boris also frames Claude Code as closer to a Unix utility than a product. It is meant to be composed into workflows the way you would pipe grep or cat, rather than wrapped in a big interface.
Claude writes most of Claude Code
Asked how much of Claude Code was written by Claude Code, Boris estimates 80 to 90 percent, with a lot of human code review on top. His default is to let Claude write the code and step in if the result is poor. He still does some work by hand, such as intricate data model refactors where he has strong opinions. He says he has not manually written a unit test in many months. He used to feel like a jerk asking for tests on someone's pull request; now he always asks, because Claude can just write the test.
What it cost, and how they thought about it
At the time, Cat put spending at around $6 per active user per day; the show notes compare that with $20 a month for Cursor. It is an average, not a cap: Boris mentions people at Anthropic racking up thousands of dollars a day with parallel automation. He called it a return-on-investment question rather than a cost question, since making an expensive engineer 50 to 70 percent more productive is worth a great deal. Asked for a productivity number, he said the team was still working on getting a credible one. Anecdotally, he put his own gain at about 2x, some colleagues at perhaps 10x, and people who only use it for commit messages at maybe 10%. Cat said the team wants to measure cycle time, such as first commit to merged pull request, and features that would not otherwise have been built.
Practical tips from the episode
The most reusable part is the advice on headless runs. Non-interactive mode is just the -p flag with a prompt in quotes. The example below is ours, modelled on the read-only lint job Boris describes:
claude -p "Check this repo for code comments that no longer match the code. Report them; do not edit any files."- Start with read-only jobs. Boris says -p works best for tasks that don't change files, such as a linter that reports without fixing, or the changelog the team generates from commit history, so permissions and runaway runs matter less.
- Pre-approve tools narrowly. There is no permission prompt in headless mode, so allowed tools are effectively pre-accepted. For jobs that write, Boris recommends passing a specific set of permissions on the command line, such as git status or git diff rather than all of bash.
- Isolate it if you can. Boris calls a Docker container, which you can snapshot and rewind, the ideal setup, while admitting most people won't do that work.
- Scale up slowly. Cat's advice is to try one case, fix the prompt, then go to 10, study the failures, and only then scale up.
- Know what slash commands are. Cat notes they are saved prompts, not tools; custom tools come through MCP. The team ran a GitHub Action that called a local lint command, checking things static linters miss, such as code that no longer matches its comments.
Cat also shows how fast prototyping changes product work: in meetings where sales or compliance asked for a feature, engineer Forrest would ask a few questions, and about 10 minutes later say it was built and he would merge it later.
What has changed since
Treat the numbers as a May 2025 snapshot. In the episode there was no separate Claude Code subscription; Boris said one had been talked about and that pay-as-you-go was the most popular option so far. As of November 2026, Anthropic's help pages say usage limits are shared across your Claude account, and that usage restored by a limit reset also applies in Claude Code. The $6 figure dates from May 2025 and should not be read as today's cost.
Our take
Worth the time for anyone building agent tools or deciding whether to automate with Claude Code. Boris's stated principle is that when the model is good, the simple thing usually works: a markdown file for memory and a summary for compaction. If you only steal one habit, make it the headless-run discipline: read-only first, narrow permissions, then scale up step by step.
Source: Claude Code: Anthropic's Agent in Your Terminal (Latent Space), Claude Code: Anthropic's CLI Agent (Latent Space on YouTube), What is a limit reset? (Claude Help Center)
This story first appeared in the claudefox daily briefing for November 9, 2026.