TubeReads

Building Claude Code with Boris Cherny

What happens when your first pull request at a top AI lab gets rejected — not because the code was bad, but because you wrote it by hand? Boris Cherny, creator and engineering lead of Claude Code, faced exactly that when he joined Anthropic. Within months, he went from handwriting code to shipping 20–30 pull requests a day, with the model writing 100% of every line. Yet the path from side project to one of the fastest-growing developer tools was anything but certain: Anthropic debated whether to keep Claude Code internal, fearing it would make them too productive. How did a research tool built to study safety in the wild become indispensable to engineers and non-engineers alike — and what does that mean for the future of software development?

Video length: 1:37:59·Published Mar 4, 2026·Video language: English
7–8 min read·22,000 spoken wordssummarized to 1,550 words (14x)·

1

Key Takeaways

1

Claude Code now writes approximately 80% of all code at Anthropic on average; Boris personally writes zero lines by hand, shipping 20–30 pull requests daily with Opus 4.5 writing every line.

2

Anthropic debated keeping Claude Code internal because of how productive it made engineers, but released it as a research preview to study safety in the wild — a decision that enabled rapid iteration and user feedback.

3

The fastest-growing use case for Claude Code is non-engineers: finance teams, sales teams, and data scientists now write code daily, signaling a shift toward multi-disciplinary generalists.

4

Code review has been restructured: Claude Code reviews every pull request in CI, catching ~80% of bugs; human review remains mandatory for safety and security, but the bottleneck has shifted from writing to verification.

5

The printing press analogy: just as scribes became writers when literacy democratized, software engineers will not disappear but transform — though Boris admits the transition is disorienting and requires letting go of attachment to handwriting code.

In a Nutshell

Claude Code is not just a productivity tool — it is a window into a future where the boundary between engineer and non-engineer dissolves, where prototyping replaces PRDs, and where the skill that matters most is no longer writing code but deciding what to build and how to verify it.


2

The First Pull Request That Got Rejected

Boris's handwritten code was turned down in favor of AI-generated output.

Adam rejected my PR and he was like, 'Actually, you should use this Clyde thing for it instead.' And I was like, 'Okay, cool.' It took me like half a day to figure out how to use this tool because you have to like pass in a bunch of flags and like use it correctly. But then it spit out a working PR. It just one-shotted it. And I think for me, this was my first fuel hi moment at Anthropic cuz I I was just, oh my god, like I didn't know the model could do this.

Boris Cherny


3

From Practical Hacker to Code Quality Lead

Boris learned to code selling Pokémon cards and cheating on math tests.

Boris's path into tech was anything but traditional. At 13, he taught himself HTML to make his eBay Pokémon card listings stand out — discovering that a blinking tag could double the sale price. In middle school, he programmed answers into his TI-83 graphing calculator, then built solvers when tests got harder, eventually dropping down to assembly to make programs run faster. Coding was never an end in itself; it was always a means to solve practical problems.

This pragmatic mindset carried through multiple startups, including a weed review site and a YC-backed medical software company where he shadowed doctors to understand why they weren't using the product. He learned that doctors had no time to boot up legacy systems, leading to pivots from web to Android to targeting different user personas. The lesson: hypothesis-driven iteration matters more than the technology stack.

At Meta, Boris spent seven years leading code quality across Instagram, Facebook, WhatsApp, and Messenger. He became one of the most prolific code authors and reviewers, automating code review by tracking repeated issues in a spreadsheet and writing lint rules whenever a pattern appeared more than three times. His work on the «Better Engineering» program — which mandated 20% of engineering time go to fixing tech debt — proved that code quality has a double-digit impact on productivity.


4

The Internal Debate: Keep It or Ship It?

Anthropic questioned whether releasing Claude Code would give away a productivity advantage.

💡

The Internal Debate: Keep It or Ship It?

When Claude Code's internal adoption went vertical — reaching nearly 100% of technical employees at Anthropic — leadership debated whether to release it publicly. The concern was simple: if it made engineers this productive, why give competitors the same advantage? The decision to ship came down to Anthropic's core mission: studying safety in the wild. Synthetic evals in a petri dish could never capture real-world risks like prompt injection, tool misuse, or adversarial prompting. By releasing Claude Code, Anthropic could observe how users actually interacted with agentic AI, iterate on safety mechanisms, and build the permission and sandboxing systems that would later become critical infrastructure.


5

How Claude Code Actually Works

🔁
Core Agent Loop
A straightforward query loop with evolving tool use. Tools are added and deleted constantly based on what works; the model decides how to use them autonomously.
🛡️
Swiss Cheese Safety
Multiple overlapping layers: model alignment, runtime classifiers, sub-agent summarization for web fetch, permission prompts, and static analysis of bash commands.
🔍
Agentic Search Wins
Early versions used local vector databases for RAG, but agentic search — essentially glob and grep wielded by the model — outperformed everything and avoided permission complexity.
Human-in-the-Loop Review
Claude reviews every PR in CI, catching ~80% of bugs. Engineers approve all changes. For enterprise customers, security and privacy are non-negotiable.

6

Boris's Zero-Handwritten-Code Workflow

Five parallel agents, plan mode, and notifications — or just the phone.

1

Start Five Parallel Checkouts Boris maintains five terminal tabs, each with a separate Git checkout. He cycles through them, starting Claude in plan mode (Shift + Tab twice) in each.

2

Iterate on the Plan The most critical step is refining the plan. With Opus 4.5 and 4.6, once the plan is solid, implementation is almost always a one-shot success.

3

Use Desktop or iOS for Overflow When terminal tabs fill up, Boris shifts to the Claude desktop app (with built-in Git worktree support) or even the iOS app, where he starts agents each morning while still in bed.

4

Tag Claude on GitHub PRs Boris uses the GitHub app integration daily. When reviewing a coworker's PR, he tags @Claude to request a lint rule, which Claude writes directly in the PR thread.

5

Ship Without Editing a Line During his December coding vacation, Boris shipped 10–20 PRs daily. Opus 4.5 wrote 100% of the code. The model introduced ~2 bugs; Boris estimates he would have introduced ~20 handwriting the same volume.


7

Key Numbers Behind Claude Code

Adoption, productivity, and the speed of building co-work.

Code Written by AI at Anthropic
~80%
On average across all engineers; Boris personally writes 0% by hand.
Internal Adoption Rate
~100%
Nearly every technical employee uses Claude Code daily; about half the sales team uses it as well.
Co-Work Build Time
10 days
From decision to ship, built entirely with Claude Code by a small team.
Bugs Caught by Claude in CI
~80%
Claude reviews every pull request automatically; human review catches the rest.
Daily PR Volume (Boris)
20–30 PRs
During his December 2024 coding vacation, all written by Opus 4.5 with zero manual edits.
Prototypes for To-Do List Feature
15–20
Built in a day and a half; Boris iterates extensively before choosing what to ship.

8

Agent Teams and Uncorrelated Context Windows

Swarms enable complex builds by delegating across fresh context windows.

CORRELATED CONTEXT
One Agent, One Window
The model performs tasks sequentially in a single context. Each new task sees everything that came before. This works for most use cases but eventually hits context limits or loses focus on complex, multi-faceted builds.
UNCORRELATED CONTEXT
Swarms with Fresh Windows
The lead agent spawns sub-agents, each with a fresh context window. Sub-agents don't see the parent's full history — only the prompt they're given. This is a form of test-time compute: throwing more tokens and agents at the problem improves results. Anthropic used swarms to build the plugins feature over a weekend, with agents creating an Asana board, splitting tasks, and implementing them autonomously.

9

The Printing Press Moment

Scribes became writers; engineers will become something new.

One metaphor I have for this moment in time is the printing press in the 1400s because there was a group of scribes that knew how to write. Some of the kings were illiterate who are employing the scribes. And if you think about what happened to the scribes, they ceased to become scribes, but now there's a category of writers and authors. These people now exist. And the reason they exist is because the market for literature just expanded a ton.

Boris Cherny


10

What Still Matters and What Doesn't

🗑️
Left Behind
Strong opinions on code style, languages, and frameworks. The model can rewrite in any stack, so endless debates are obsolete.
🔬
Still Critical
Being methodical and hypothesis-driven, especially for debugging and product discovery. The model helps, but humans still need the skill — for now.
🌐
More Valuable
Curiosity and working beyond your swim lane. The next billion-dollar product may come from one person who thinks across engineering, product, design, and business.
Newly Rewarded
Short attention span and rapid context switching. Managing five parallel Claude agents is the new deep work.
🔄
Adaptability
The only certainty is that the next model will change everything again. Engineers must stay open to revising workflows and discarding ideas that worked last quarter.

11

Securities Mentioned

METAMeta Platforms, Inc.

12

People

Boris Cherny
Creator and Engineering Lead, Claude Code; Former Code Quality Lead at Meta
guest
Gergely Orosz
Host, The Pragmatic Engineer
host
Adam Wolf
Boris's Ramp-Up Buddy at Anthropic
mentioned
Fiona Fun
Manager, Claude Code Team
mentioned
Dario Amodei
CEO, Anthropic
mentioned
Andrej Karpathy
AI Researcher
mentioned
Ben Mann
Founder, Anthropic; Started Labs Team
mentioned

Glossary
Agentic AIAI that can autonomously use tools, execute multi-step plans, and make decisions without requiring conversational back-and-forth with a human.
MCP (Model Context Protocol)A universal protocol that allows AI applications to access external tools and data sources, enabling agents to interact with systems like SonarQube or IDEs.
Uncorrelated Context WindowsMultiple AI agents working in parallel, each with a fresh context window that doesn't inherit the parent agent's full history, enabling more parallelism and test-time compute.
Prompt InjectionA security risk where malicious instructions embedded in data (e.g., a webpage) trick the AI into performing unintended actions, like deleting files.
Git WorktreeA Git feature that allows multiple working directories for the same repository, enabling parallel work on different branches without needing separate clones.

Disclaimer: This is an AI-generated summary of a YouTube video for educational and reference purposes. It does not constitute investment, financial, or legal advice. Always verify information with original sources before making any decisions. TubeReads is not affiliated with the content creator.