Last week, we hosted Boris Cherny, creator and Head of Claude Code, at the Community Meetup in Ankara sponsored by Anthropic. Other speakers included Fatih C. Akyon, Emre Yılmaz, Eray Yapağcı, and myself. Here are the key insights from Boris.

Key Insights from Boris Cherny

100% AI-Generated Workflow

  • Adoption Rate: Boris transitioned from ~10% AI assistance a year ago to 100% with Opus 4.5.
  • Volume: He shipped over 300 pull requests in the last month without writing manual code.
  • Orchestration: Runs multiple instances simultaneously (3 terminal agents, 3 to 5 web agents, plus mobile agents), acting as a manager rather than a contributor.

Operational Strategy

  • “Boring” Efficiency: Workflow relies heavily on Plan Mode followed by auto-accepting edits.
  • Friction Reduction: Uses a comprehensive allowlist in “settings.json” to suppress repetitive permission prompts.
  • Steering vs. Security: As models gain intelligence (Opus 4.5), manual “steering” (telling the model what not to do) is becoming obsolete; the model instinctively understands intent.

The End of Reading Code

  • Prediction: By next year, engineers will likely stop reading source code entirely.
  • New Abstraction Layer: Work will shift from manipulating source code to manipulating agent transcripts.
  • Self-Verification: Opus 4.5 is the first model trusted to verify its own work, autonomously writing unit tests and using Puppeteer for browser-based testing without human intervention.

The Changing Skill Stack

  • New Assets: A short attention span is now a superpower, allowing engineers to rapid-fire context switch between multiple active agents.
  • Enduring Skills: Syntax knowledge is fading, but systems architecture, language design, and maintainability remain critical. You must understand the layer below where you work to effectively manage the bot.

Advice for Builders & Enterprises

  • The 1-Person Unicorn: We are approaching the era of the first billion-dollar company built by a single person. Claude can now handle frontend, backend, and mobile simultaneously.
  • Enterprise Context: For massive repos, drop a “CLAUDE.md” file not just in the root, but in every subdirectory (services, frameworks) to provide localized context.
  • Share Intelligence: Do not keep personal prompts or agent configs private; check them into the codebase so the entire engineering team benefits from the agent’s tuned behavior.

The Power of the SDK (Headless Agents)

  • Programmatic Agents: The Cloud Agent SDK allows for headless, programmatic agent loops.
  • Case Study: An engineer built the Go SDK autonomously by setting up a cron job that triggered Claude to file its own GitHub issues, spin up sub-agents to solve them, write tests, and generate the code without human prompting.

Advice to Students

  • Build a Clone: Build your own version of Claude Code from scratch to fundamentally understand the UX and LLM mechanics.
  • Start Now: Stop studying just to code; start a company immediately using these tools to build entire products.

Thanks to the Bilkent AI Society, Anthropic, and everyone who made this event possible.