Inside Claude Code With Its Creator Boris Cherny (AI column)

Interesting. interview Anthropic's Boris on Y Combinator's Lightcone podcast. The discussion revolved around Claude Code and how Boris created one of the most successful AI tools in the form of a simple terminal product, and what happens next. ​ When it comes to key insights, they are

🚀 Build under the model through 6 months In Anthropic, the principle is not to optimize the product for the current model, but to think about what it will be in about six months. Any complex "scaffolding" around the model often gives +10–20% and is completely reset by the next release, so it is better to have a minimum layer of binding and wait for the upgrade of the model.

🎉 The birth of Claude Code as a side effect Boris just wanted to learn how to use the Anthropic API and wrote a terminal chat client without a UI. Added bash-tool, asked the model “to know what music I listen to”, she generated AppleScript and climbed into the player – for him it was the first “AGI-moment”: the model “greatly wants to use tools”.

Why the terminal and why it's stuck The terminal was not chosen out of ideology, but as the cheapest way to make a prototype. Inside Anthropic, people started using it "before it was ready," because: Convenient to automate git, bash, Kubernetes, routine DevOps-carrying The product feels like a “game” rather than a heavy IDE plugin. From this grew the principle: follow latent demand (latent demand) Look at what people are already trying to do and make it easier, rather than imposing a new stream of work. This is a top-end product approach for building internal development tools.

✍️ CLAUDE.md / CLAUDE.md as an operational context Boris has a personal CLAUDE.md - just two lines: turn on automerge on PR and post PR in the command channel for a quick review. The rest of the “politics” lives in the repo-local CLAUDE.md, which the entire team rules several times a week. He advises that if a file has swelled to thousands of tokens, just delete and reassemble – with each model you need fewer instructions.

🔈 Verbosity and UX terminal The team is constantly tuning the details: they tried to hide the bash withdrawal, the employees rebelled – it is needed for debag (Kubernetes, complex teams). Hidden detailed file reading/search logs, but GitHub users added verbose mode in config.

💪 How he works with Claude Code himself

• 80The percentage of sessions starts in plan mode: first plan, then execution. Parallels work: several tabs in the terminal and desktop application, each starts with a plan. For complex tasks, explicitly asks for multiple subagents (3, 5, 10) explore the problem in parallel and then combine the conclusion.

🔮 The future of plan-mode and agents Plan mode is just one extra phrase in the "please don't write the code" prompt. He believes that the lifetime of the explicit plan mode is limited: with the growth of the model’s abilities, it will itself enter the “planning mode”, and then it will be possible to give the task in one shot. Already, Claude Code sometimes includes a plan-mode when it would be natural for a person.

Automation of the entire engineering chain Inside Anthropic Claude Agents (via Claude Agents SDK) Automate: code review, security review, triage and labeling tasks, the path of the feature to production. Plugins feature for Claude Code was completely written by a swarm of agents over the weekend: one agent got speculated, created tasks at Asana, spawned sub-agents, they raised PR-s.

🤖 Profile of the ideal engineer/founder in the LLM era The main thing is the scientific approach, thinking from first-principles and a willingness to admit mistakes; many experienced engineers are stuck in old patterns. Boris appreciates either hyper-specialists (as a bun team / people obsessed with devtools, rantimes)Or hyper-generalists who cross product, engineer, center, business. Selection of candidates - including behavior with agents: can you understand from the transcript of the session that a person can think systematically, use a plan, logs, adjust the model?

#AI #Engineering #Software #Management #Leadership #Startup #LLM #ML #Architecture