Anthropic had to use Claude Code to make Claude Code

In March 2026, when Claude Code source code was leaked, everyone mocked Anthropic. Famously, they used Claude Code to write 100% of Claude Code. Infamously, the code was terrible. But I thought that they were right all along to use Claude Code to write 100% of Claude Code.

In fact, I think that building their AI platform while using this very platform has always been mandatory for any AI company.

First, let’s look at the practicality of building a coding tool. Sure, Claude Code tooling was not ready to replace a software engineer drop-in in 2025 and early 2026, definitely not with Sonnet 4.5. But. Anthropic claimed that their AI was good enough to write code. Here we have a contradiction: Anthropic were claiming Claude Code was great, but their customers said it was not good enough. What was the best possible way for Anthropic to find out and fix the pain points?

A well-known solution is “dogfooding” (which is a terrible name, come on). In simple terms, it means “use your product yourself”. Use Claude Code and the underlying model for your critical work and suffer the consequences of it failing. This was the best way for Anthropic engineers to find the limitations of Claude Code and improve it, and they followed it. With the power of hindsight from July 2026, we now know that it has paid off. Today’s Claude Code tooling is brilliant, and it harnesses and assists the latest models excellently. The huge manual setup and extra prompts I had to build around the Sonnet 3.7 and Claude Code in 2025 were no longer needed. Today’s Code works better than anything I have built in 2025.

Dogfooding alone justifies using Claude Code to develop Claude Code. But,

For an AI company, the stakes are even higher

In 2019, when asked how OpenAI would generate revenue, Sam Altman famously replied:

“We have no idea how we may one day generate revenue. We have made a soft promise to investors that once we’ve built this sort of generally intelligent system, […] we will ask it to figure out a way to generate an investment return for you.”

Remember, the year is 2019. GPT-3 does not exist yet. ChatGPT will not exist for another 3 years. Claude - what is Claude? So, naturally, people in the audience laughed. I laughed. And you should too - even Sam acknowledges how much of a “Silicon Valley” this answer is.

Art is defined by getting better the more you look at it, and thus this answer is art. Because the more you think about it, the more it becomes obvious that it is the only right answer! If you truly believe that you can build something to do the intellectual work, and people will pay for it, then it must be good enough to do the said work - and also in terms of business model analysis! So if you believe that you are not a scam but actually building the most capable AI, then how can answering “we will ask AI” not be the only right answer?

Today, we all use LLMs to do our work, so it seems obvious. In 2019, it felt like a joke. In early 2026, the Anthropic team using Claude Code to write Claude Code felt like a joke again. Yet again,

if they truly believed that Claude Code is so good their customers should pay for it for saving money developing software, they had no such option as to not use it themselves.

Not to mention that an AI platform company not using its own AI is sending a clear signal that they are a scam. Of course, they must use their own product.

P.S.

I wrote this in April 2026, right after the code leaked. Back then, I also wrote:

Using Claude Code to make Claude Code, short-term, people laugh at you, and your code might have degraded temporarily. But only by failing can one learn what to fix -> and that is why I have a strong feeling that Claude Code and Opus models may become greater in 6-12 months than anything we could have handwritten without their own help. In 2027, we will look back at 2026 with the power of hindsight, and confidently claim “of course it was obviously a great idea Anthropic used early Code to develop Code, and I have always been saying so”.

Turns out, I was too pessimistic. It has been only 3 months since the source code leak, and Claude Code has already evolved into a mature tool. This idea already feels obvious. I really shouldn’t keep my writing about AI in the drafts for too long - the ground is shifting too fast.