The Secrets of Claude Code From the Engineers Who Built It
Boris and Cat, the team behind Claude Code, explain how a terminal tool grew out of an internal prototype after an engineer was told to use it instead of hand-writing code. They cover why giving the model bash and a few tools worked better than building a custom toolkit.
- Claude Code began as an internal research tool called Clyde. A new engineer's hand-written PR got rejected with "use Clyde," and the one-shot result convinced the team that an IDE might not be necessary.
- The turning point was giving the model bash. It started writing AppleScript and automating things on its own, which the team had not expected.
- Anthropic "dogfoods" the product: roughly 70 to 80 percent of technical staff use it daily, and internal feedback arrives constantly, so features get shipped or pulled within days.
- The team has been removing tools over time. The LS tool was dropped once bash permissions could enforce directory restrictions, which leaves less clutter in the model's context.
Boris and Cat, two members of the Claude Code team at Anthropic, walk through how the tool came about. It wasn't planned. Cat joined Anthropic and, as a new hire, hand-wrote his first pull request in an editor. His manager, Adam Wolf, rejected it and told him to use Clyde, an internal CLI that Anthropic engineers were already hacking on. That tool was a research project: slow to start, heavy Python, lots of indexing. Cat said Clyde one-shot the task on Sonnet 3.5, though the harness was so old that it took around five minutes and he still had to fix the output. The real turning point came when Cat was prototyping against the Anthropic API and built a small terminal chat app to avoid writing a UI. Giving the model tools, especially Bash, was the revelation. The model reached for Bash on its own, wrote AppleScript to automate tasks in response to questions, and generally behaved as if it wanted to use tools. Cat says the "oh, this actually works" moment came with Sonnet 4 and Opus 4.
The design choice that makes Claude Code distinctive is that it leans on Bash rather than a large set of bespoke wrappers like find-file or open-file, which is what many agent architectures start with. Claude Code does still have a dozen or so tools, and the team adds and removes them most weeks. Dedicated search exists for two reasons: a nicer display for the human in the loop, and permissions. Enforcing a rule like "this file cannot be read" is easier when there's a specific tool to gate. Recently they removed the LS tool, because they built a way to enforce directory-level permissions inside Bash. Fewer tools means less choice for the model and less clutter in context.
On the business side, the speakers describe a split of responsibilities. Boris sets technical direction and has driven much of the product vision. Cat's role is more supporting: making sure pricing and packaging land with users and shepherding features through launch. The conversation also touches on power-user behavior, with some Anthropic staff spending over $1,000 a month in credits. Their framing borrows from YC and the old product idea of latent demand: build something hackable and open-ended enough that people use it for purposes you didn't design for, then build toward that demand. The host also asked whether the CLI is the final form factor, and the answers in the excerpt are cut off before that question is resolved.