Back to Videos

Live Coding Session with Claude Code and Boris

Channel Anthropic
Date May 6, 2026
Duration 32 min
Tags Claude Code, Live Demo, Developer Workflow, Boris Cherny, Anthropic
TL;DR

Boris Cherny, creator of Claude Code, shows his actual day-to-day workflow in an unfiltered live session. Covers real project work, how he structures prompts, when he lets Claude run autonomously versus when he intervenes, and the parts of his workflow that might surprise people who use Claude Code differently.

Key Takeaways

Summary

The Setup

Boris opens with his typical session initialization: loading a CLAUDE.md with project-specific context, using tool search rather than a fixed toolset, and starting with a short planning prompt that gets Claude to describe its approach before acting. He does this for every non-trivial task — even after years of use.

Real Work, Unfiltered

The session is built around actual work — not a prepared demo. Boris picks up where he left off on a real project, and the rough edges are visible: Claude misunderstands an instruction, he corrects it; a tool call fails, he adjusts; context gets long, he compacts. The value is seeing how he handles these moments, not just the happy path.

Parallel Agents in Practice

Mid-session, Boris demonstrates spawning a review agent to check his recent changes while he continues working. The review agent runs in the background, returns with findings, and he addresses the high-priority ones without context-switching out of his main flow. This is the "team" model of Claude Code use in practice.

When to Intervene

Boris's rule of thumb: let Claude run for longer than you're comfortable with, then verify the result structurally (does the shape look right?) before reading it in detail. He rarely reads Claude's output line by line; he runs tests, checks diffs, and trusts the verification layer rather than the prose.

CLAUDE.md Philosophy

His CLAUDE.md files are short and non-obvious — not a list of language preferences but a set of constraints and invariants that Claude would otherwise guess wrong. He recommends writing them after you've caught Claude making the same mistake twice, not before you start the project.

Notable Quotes

"I built this thing and I still learn something about how to use it every week. That should tell you something about how much is possible."

"The best CLAUDE.md entries are the ones that sound obvious once you read them but you'd never think to write them until Claude burned you."

References