Inside Anthropic: How Claude Code Engineers Ship with 2-Engineer Teams and

What Changed — Inside Anthropic's Engineering Culture

A deep dive into how Anthropic builds software reveals a radical shift in engineering practices. The Pragmatic Engineer visited Anthropic's San Francisco office and interviewed four key engineers: Katelyn Lesse (Head of Engineering for Claude Platform), Jarred Sumner (creator of Bun, now on Bun and Claude Code), Thariq Shihipar (Claude Code engineering and education), and David Hershey (Applied AI, working with Cursor, Cognition, Perplexity).

Key changes:

• Two-engineer teams: Max two engineers per project. No more large squads.
• AI-driven code review: AI handles an increasing share of code review and testing.
• Prototyping is fluid: Design is ongoing, not upfront. Teams work on more projects.
• Massive rewrites in days: A 500K+ line Bun rewrite to Rust took 11 days using Fable and $165K in tokens.

But some things stay the same: two-pizza teams are alive, planning matters for complex projects, PRDs are still used, context switching remains a challenge, and the ratio of coding to testing time hasn't changed much.

What It Means For You — Concrete Impact on Daily Claude Code Usage

1. Embrace the Two-Engineer Rule

Anthropic's max-2-engineers-per-project approach forces focus. For Claude Code users, this means:

• Use CLAUDE.md to codify project context so a single engineer (or pair) can onboard fast.
• Leverage Claude Code's Plan mode (introduced July 28, 2026) as a safety rail for cross-file refactors—no need for a 10-person planning meeting.
• Keep projects small. If a task needs more than 2 engineers, break it into smaller projects.

2. AI Code Review Is the New Normal

Anthropic reports that AI now handles a significant portion of code review and testing. For you:

• Use Claude Code's review capabilities on every PR. Run claude code review to catch issues before human review.
• Automate testing with AI. Anthropic's approach: implementation is fast, verification takes longer. Let Claude Code generate and run tests.
• Don't skip human review—AI handles the grunt work, humans focus on architecture and edge cases.

3. Massive Rewrites Are Now Feasible

Jarred Sumner's 11-day rewrite of Bun (500K+ lines) from JavaScript to Rust using Fable and $165K in tokens is a proof point. For your projects:

• Consider rewrites you previously dismissed. With Claude Code and $100-200K in tokens, a year-long migration can become a week-long task.
• Use Fable for language-to-language transpilation. It's not just for Bun—try it on your legacy codebase.
• Budget for tokens, not engineering months. $165K seems large, but compare to a 12-month team salary.

4. Design Is Ongoing, Not Upfront

Anthropic teams don't spend weeks on design docs. They prototype fast and iterate. For Claude Code users:

• Start coding immediately. Use Claude Code to generate a prototype from a sentence description.
• Iterate in Claude Code's chat. Don't finalize architecture before writing code—let the AI help you discover it.
• Use skills (Claude Code's new skill packs in ~/.claude/skills/) to share reusable design patterns across your team.

Try It Now — Commands, Config, and Workflow Changes

Adopt the Two-Engineer Rule

# In your CLAUDE.md, add project scope to keep teams small
# ~/my-project/CLAUDE.md
# This project is scoped to 2 engineers max.
# Use Plan mode for any cross-file changes.

Automate AI Code Review

# Add a pre-commit hook that runs Claude Code review
# .git/hooks/pre-commit
claude code review --staged --output-format json > review.json
if grep -q "\"critical\"" review.json; then
  echo "Critical issues found. Commit blocked."
  exit 1
fi

Run a Massive Rewrite Experiment

# Prototype: rewrite a small module to Rust using Fable
# WARNING: This costs tokens. Start small.
fable your-module.js --target rust | claude code --stdin --output rewritten.rs

Embrace Ongoing Design

# Instead of writing a design doc, start with a prototype
claude code "I need a REST API for user management. Start with a single endpoint and expand."

Why This Matters for Claude Code Users

Anthropic's own engineering practices validate what Claude Code enables: small teams shipping faster with AI as a core collaborator. The key takeaways:

1. Small teams + AI > large teams. Two engineers with Claude Code can match a 10-person team.
2. AI code review is production-ready. Use it now.
3. Massive rewrites are no longer impractical. Budget tokens, not months.
4. Design through code. Prototype first, plan later.

These aren't theoretical—they're how Anthropic builds the tools you use.

Source: newsletter.pragmaticengineer.com