Y Combinator

Boris Cherny: How Tool Use Turned a CLI Into Claude Code

By betting on bash tools and a minimal terminal loop; CLAUDE.md emerged from organic user demand, and subagents now parallelize repo exploration.

BCBoris ChernyguestGTGarry TanhostJFJared Friedmanhost

Feb 16, 2026·50m·Watch on YouTube ↗

🙋 FREQUENTLY ASKED QUESTIONS

Direct answers grounded in the episode transcript. Tap any timestamp to verify against the source.

1. 1

   Why is Claude Code still in the terminal?

   Claude Code stayed in the terminal because the CLI kept proving useful while models were changing fast. Boris says the terminal started as an accident: after the first prototype, he gave it to his team for dogfooding, and Robert, an engineer sitting across from him, was already using it to code the next day. During the external launch review in late 2024, Dario asked whether engineers were being forced to use it because the internal daily active usage chart looked vertical. Boris said no one was mandating it. He had only posted about it, and engineers spread it by word of mouth. The CLI was also the cheapest starting point, and the team felt there was no UI they could build that would still be relevant in six months because the model was improving so quickly.

   ▸ 6:33 in transcript
2. 2

   What goes in Boris Cherny's CLAUDE.md file?

   Boris Cherny keeps his personal CLAUDE.md file intentionally tiny. He says it has two lines: one tells Claude to enable auto-merge whenever he opens a PR, and the other tells Claude to post his PR in an internal team stamps channel so someone can stamp it and unblock him. Other instructions live in the CLAUDE.md checked into the code base, which the whole team updates multiple times a week. When a preventable mistake appears in a PR, Boris will tag Claude and ask it to add the lesson to the shared CLAUDE.md. He warns against over-engineering the file. If it grows to thousands of tokens, his recommendation is to delete it, start fresh, and add back only the instructions needed when the model gets off track.

   ▸ 9:01 in transcript
3. 3

   How do Claude Code subagents use uncorrelated context windows?

   Claude Code subagents use separate context windows so parallel agents can explore a problem independently. Boris describes agent topologies as a new field for configuring agents, and names uncorrelated context windows as one sub-idea. The setup is multiple agents with fresh context windows that are not polluted by each other's context or by their own previous context. Throwing more context at a problem works like test-time compute, so the system can gain capability when the agents can communicate in the right topology. Claude Teams is one version of that idea. Boris says Anthropic's plugins feature was built by a swarm over a weekend: an engineer gave Claude a spec and an Asana board, then Claude created tickets, spawned agents, and had them pick up tasks.

   ▸ 22:00 in transcript
4. 4

   How does Claude Code plan mode actually work?

   Claude Code plan mode is a simple prompt behavior that tells Claude to think before coding. Boris says there is no big secret to it: it adds a sentence to the prompt like, "Please don't code." The feature came from watching users ask Claude to plan something, write a spec, or talk through an idea without starting implementation yet. He wrote the first version on a Sunday night after reading GitHub issues and an internal Slack feedback channel, then shipped it Monday morning. Boris still uses plan mode heavily, starting about eighty percent of his sessions that way. Once the plan is good, he has Claude execute, and newer models stay on track much more reliably than they used to.

   ▸ 25:26 in transcript
5. 5

   What does Boris Cherny mean by building for the model six months ahead?

   Building for the model six months ahead means designing around capabilities that are improving, not static. Boris says LLM founders should use the current model, feel out the boundary of what it can do, and then build for the version they expect a few months later. Otherwise, a team may find product-market fit for today's model and get leapfrogged when a new model changes the tradeoffs. He connects this to Rich Sutton's The Bitter Lesson, which he says is framed in the Claude Code area. The lesson inside Claude Code is to never bet against the model. Product code around the model, which the team calls scaffolding, can add perhaps ten or twenty percent in some domain, but it can become tech debt when the next model does the same thing directly.

   ▸ 37:43 in transcript

Answers are AI-generated from the transcript and may contain errors. Tap a question to verify against the source.