techlifeadventuresVol. 03 · Nov 2026
Claude Code Writes Claude Code: What That Claim Means
·5 min read·AI & Machine Learning

Claude Code Writes Claude Code: What That Claim Means

Claude Code's creator said Claude wrote all of his code for a month. What that claim did and didn't mean, set against my own six months of real numbers.

Claude Code Writes Claude Code: What That Claim Means

I first wrote this post in January 2026, when a few posts from well-known engineers made it sound as if programming had changed overnight. I've rewritten it in November 2026. The original leaned on quotes and hype. Since then I've spent six months running my own engineering work through Claude Code and measured what happened, so I can now check the claims against something real.

What was actually said

Boris Cherny, who created Claude Code at Anthropic and leads the team, posted in late December 2025 that he hadn't opened an IDE in the previous month. His numbers for that month:

MetricValue
Pull requests259
Commits497
Lines addedabout 40,000
Lines removedabout 38,000
Written by Claude Code (running Opus 4.5)all of it
My January version of this post said "100% of Claude Code's codebase was written by Claude Code." That was wrong, and the mistake matters. The figures were Cherny's own contributions for one month, not the whole team's or the whole codebase. People also asked fair questions under his post: how many of those PRs reached production, and what it means when the tool is free to the person using it.

Around the same time (26 December 2025), Andrej Karpathy posted that he had never felt so far behind as a programmer. He described the profession as being refactored, with programmers writing a smaller share of the code themselves. It was widely shared because of who said it, not because it contained data.

Peter Steinberger, the developer behind the OpenClaw agent, became known for saying he ships code he doesn't read. In interviews he made clear he relies on tests, linting and a review of the architecture instead of line-by-line reading. That caveat usually got lost when the line was shared.

I've dropped several other quotes that appeared in the original version (from Shopify's CEO, Eric Schmidt, an xAI co-founder and a course creator). I couldn't trace them back to a primary source with confidence, and a post about checking claims shouldn't repeat claims it can't check.

What my own numbers say

Between April and October 2026 I moved my day-to-day work as an engineering leader onto Claude Code agents. I wrote it up in full in Moving an engineering team to agent-driven delivery: six months of real numbers. The short version, counting my own PRs only:

Metric (per month)BeforeAfter
Merged pull requests11.6113.6
Median time from first commit to merge2.0 hours0.7 hours
Test cases addedabout 470about 1,350
Reverts stayed rare: three to four in twelve months. So yes, the order of magnitude Cherny described is reachable outside Anthropic. I got to about 114 merged PRs a month against his 259. But three things in my data change how I read his claim:
  1. The PR count overstates the change. My PRs grew about 10x while lines added grew about 4x. Agent PRs are smaller. "10x more PRs" is not "10x more software."
  2. Other things changed at the same time. My role grew and the number of repositories I worked in doubled. I can't credit all of the increase to the agent, and I doubt anyone posting a single month's numbers can either.
  3. Speed came with new kinds of failure. An agent invented a ticket ID that ended up in the changelogs of two published package versions, which can't be edited. A scheduled agent posted my private morning brief to a team channel four times. Parallel sessions sharing one checkout overwrote each other's work. None of that shows up in a PR count.

What I'd take from it

The useful lesson from early 2026 wasn't "humans are done writing code." It was narrower than that, and more practical.

  • The work moves to specifying and checking. Writing clear instructions and reviewing what comes back became most of my job. The prompt engineering guide covers the instruction side.
  • Not reading the code only works with something else doing the checking. Steinberger relies on tests. I rely on tests, a read-only architecture reviewer agent and a rule that any ID or number an agent cites must be looked up first. On this site, I skipped that checking and shipped three broken tools.
  • Route work by model tier. In a two-week audit, 86% of my sub-agent work ran on mid-tier or small models. The top model is for judgement, not for searching files.
  • Measure before you believe a multiplier, including your own. I treat 4x to 10x as the fair range for my own work, depending on which number you pick, and I explain why in the case study.

Where things stand in November 2026

The models have moved on several times since Opus 4.5. As of November 2026, Claude Code defaults to Claude Opus 5.5, with Claude Fable 5.1 as the most capable tier and Sonnet 5.5 and Haiku 5.5 for lighter work. Run /model in Claude Code to see what your plan offers, because the lineup keeps changing.

If you want to try it yourself, my Claude Code tutorial for beginners covers installation, plan mode, CLAUDE.md and the commands I use daily. If you want to see what building a whole project this way looked like, read how I built this site with AI.


Related Reading:

Enjoying this article?

Get posts like this in your inbox. No spam, unsubscribe anytime.

Share this article
Vinod Kurien Alex

Vinod Kurien Alex

Engineering Manager with 20+ years in software. Writing about AI, careers, and the Indian tech industry.

Related Articles

© 2026 TechLife AdventuresBuilt with care · v3.2.1