WhoBuilt

Inside Claude Code With Its Creator Boris Cherny

Feb 17, 2026· 50 min· 253K views·Watch on YouTube
#Coding Agents#Product Management#AI Benchmarks#Developer Tools#Customer Feedback#AI Alignment

Speaker

Boris ChernyCreator and Head of Claude Code at Anthropic

Boris Cherny, creator and engineer of Claude Code at Anthropic, talks with Y Combinator’s The Lightcone hosts about building for improving models, product iteration, and how AI is changing software work.

A terminal prototype became Claude Code

Cherny says Claude Code began as a small terminal chat app he made to learn Anthropic’s API, then grew as he experimented with tool use. After giving the model Bash, he was surprised when Sonnet 3.5 used AppleScript to identify the music playing on his Mac. He shared the prototype with his team within days, and engineers began using it before it was ready.

Build from user behavior and keep scaffolding small

Early users applied Claude Code to Git, Bash, Kubernetes, and unit tests, and started writing Markdown instructions that led to CLAUDE.md. Cherny calls this latent demand: make tasks people already try to do easier. He recommends starting with minimal instructions, adding guidance when the model gets something wrong, and expecting to revise scaffolding as models improve.

Adapt workflows as models get more capable

Cherny describes iterating on how much tool output Claude Code shows: employees wanted Bash output visible, while file reads and searches could later be summarized as the model became more reliable. He says plan mode grew from users asking Claude to work through an idea without coding; his own sessions often start there. As models improve, he expects people to need less babysitting and some workflow features to change or disappear.

Use parallel agents for larger tasks

Cherny describes Claude Code teams as one way to coordinate agents with separate context windows. For a plugins feature, an engineer gave Claude a spec and an Asana board; Claude created tickets and launched agents to work on them over a weekend. Cherny also says the team uses the Claude Agent SDK for tasks such as code review, security review, issue labeling, and production workflows.

Build for the next model and broaden engineering work

Cherny advises founders to explore what current models struggle with and build for capabilities they expect within about six months. He says Claude Code is repeatedly rewritten, with tools added or removed every few weeks and little code remaining unchanged over six months. At Anthropic, he says productivity per engineer has risen 150% since Claude Code launched, and predicts engineers will spend more time on specs, users, and product work alongside coding.

Takeaways

  • Watch what users already attempt, then make that workflow easier.
  • Keep model instructions minimal and add them when a repeated failure shows they are needed.
  • Revisit product scaffolding as model capabilities change; yesterday’s workaround may become unnecessary.
  • Use multiple agents with separate contexts for difficult research or implementation tasks.
  • Build toward likely model improvements while validating ideas with users.

Mentioned in this talk

AnthropicClaude CodeClaudeVS CodeClaude Agent SDKAsanaClaude OpusCo-workplan modeMCPBashTypeScriptGitVimBitter Lesson

Search inside this talk

What was claimed 11

Related talks