From Hand‑Written Code to 30 Daily PRs: Lessons from the Creator of Claude Code

Hello, I’m Tony Bai.

Imagine joining a top‑tier AI lab, submitting your first pull request, and having it rejected not because the code is bad but because it was written by hand. This mirrors the real experience of Boris Cherny when he joined Anthropic.

Claude Code’s Birth: Don’t Put AI in a Box

Claude Code originated from an internal prototype called “Clyde.” Boris initially treated the model as a component with strict input‑output interfaces, a common mistake. He soon realized that the model should be seen as an independent entity with tools, allowing it to run programs directly.

“Don’t try to force it into a box or make it act in a specific way. Treat the model as an independent entity, give it tools, and let it run programs on its own.”

This “Bitter Lesson” philosophy gave the model permission to execute Bash commands and read/write the file system. In one early demo, Boris gave the model a Bash tool and asked, “What music am I listening to?” The model wrote an AppleScript that used sed to query the local music player and returned the answer, impressing Boris with a glimpse of AGI‑like capability.

From Hand‑Writing to “Conducting”: Parallel Agents Workflow

As a former high‑output Meta engineer, Boris now produces 20‑30 PRs per day, ranging from a few lines to thousands. He achieves this with massive parallel agents:

Open multiple terminal tabs (e.g., five) in tmux, each checking out an independent codebase, optionally using Git Worktree.

In each tab, start Claude Code in “Plan Mode” by pressing Shift+Tab twice and describe the task to the agent.

Poll‑conduct: as soon as the first agent begins thinking and executing, switch to the next tab and launch another agent, looping continuously.

Validate and deliver: when an agent signals task completion, switch back to verify the result.

In this mode Boris is no longer a “typist” but a “conductor,” focusing on defining type signatures for business logic and validating model output rather than hand‑coding implementations.

When AI Writes 80% of the Code, How Do You Review?

At Anthropic, up to 80% of code is generated by Claude Code. Quality is ensured through an “AI reviews AI” pipeline supplemented by a human safety net:

Agent self‑testing: Claude Code automatically writes and runs tests locally; if engineers modify Claude Code itself, the agent spawns a subprocess for end‑to‑end testing.

AI pre‑review (Best of N): During CI/CD each PR is first examined by Claude Code. To mitigate nondeterminism and hallucinations, multiple agents run in parallel and a deduplication agent aggregates results, catching roughly 80% of low‑level bugs.

Dynamic lint rules: When a recurring issue is detected, Boris asks Claude to generate a lint rule on the spot, preventing the problem at its source.

Human final approval: Despite high automation, every PR still requires a real engineer to perform a second‑round review and give final sign‑off for production‑grade code.

We Are Like 15th‑Century Scribes

Even AI pioneers like Andrej Karpathy feel “more behind than ever.” Boris likens the rise of AI coding tools to the invention of the printing press. Before printing, literacy and copying were privileges of a few; after printing, books became cheap and abundant, spawning new professions.

“Our software engineers are like those scribes, and the business side (CEO/PM) is like the illiterate king.”

Today Claude Code is the “printing press” of software engineering, lowering the barrier so that non‑technical staff—designers, finance people, even CEOs—can directly turn ideas into software, exploding both output and application scope.

Engineers’ New Survival Rules: What Skills Depreciate, What Appreciate

Rapidly depreciating skills:

Religious zeal for specific languages or frameworks (e.g., endless debates over React vs. Vue or Go vs. Rust). The model can rewrite code in any language within minutes.

Obsessing over syntax details; repetitive boilerplate will soon be auto‑generated.

Increasingly valuable core abilities:

Systematic, hypothesis‑driven thinking for complex debugging.

Cross‑disciplinary curiosity; full‑stack fluency plus business, design, or finance knowledge enables a one‑person team to build billion‑dollar products with AI assistance.

High‑frequency context‑switching (ADHD‑style work): managing multiple AI agents requires rapid shifting between high‑level contexts rather than deep, prolonged coding sessions.

ADHD‑style work is a flexible, highly distributed attention approach that breaks large tasks into small, manageable goals, allowing rapid iteration and frequent context changes.

Conclusion: Shed Arrogance, Embrace Change

Boris admits he often feels the pace of model evolution is overwhelming; ideas that failed months ago can succeed instantly with a new model. In this era, intellectual humility outweighs past experience. Accept that AI can write faster and sometimes better, and learn to conduct the AI‑driven symphony rather than clinging to manual coding.

The future belongs not to those who reject AI, but to those who know how to harness it to build the next generation of software.