Anatomy of the .claude/ Folder
DAILYDODEOFDS.COM
Last week I shared a Claude Code Cheat Sheet that I am finding very useful. Claude Code, like many chat interfaces, has a discoverability issue (something I wrote about just over a year ago), it is far from obvious what it is capable of, and in the AI-age we are even less likely to read the manual.
This blog post is a great complement to the Cheat Sheet, taking a really deep dive into how your Claude Code configuration is encoded via the various files and folder structures. I’m guilty of treating the .claude folder as a black box, just post this post describes. And I know that it has slowed me down in the past, as I frustratingly ask Claude Code to ‘fix itself’ with mixed results.
As with most things, building up your own knowledge, rather than just leaning on AI, leads to a better solution in the long run. I am going to take a more active interest in what lurks within the .claude folder.
Copilot Edited an Ad Into My PR
ZACHMANSON.COM
In this (brief) post, Zach drew attention to an advertisement at the bottom of a pull request opened by GitHub Copilot.
Apparently this has been going on for a while, with an astonishing 1.5 million pull requests featuring messages such as this:
“📍 Connect Copilot coding agent with Jira, Azure Boards or Linear to delegate work to Copilot in one click without leaving your project management tool.”
Interestingly, these were originally intended to be tips, a relatively innocuous feature. However, over time, the tips have morphed into product recommendations … Ads.
The GitHub team rushed to turn this feature off on the same day that it was highlighted by Zach’s post. Another bad PR week for GitHub, it was only last week that they published a widely derided post about their change in data policy - where they now [train on your data unless you explicitly opt out]((https://github.blog/news-insights/company-news/updates-to-github-copilot-interaction-data-usage-policy/), presenting this as a ‘feature’.
Features, tips - someone at GitHub needs a dictionary!
The Claude Code Source Leak
ALEX000KIM.COM
And the big news this week is … the sourcecode for Claude Code was leaked!
The leak occurred as part of their standard distribution approach via npm. Unfortunately due to some internal error, they also shipped the a .map file with their latest release. This is a sourcemap file that is typically used for debugging released (minified and obfuscated) artefacts, by providing a map to a non-obfuscated version of the code. In other words, the sourcemap provides the original source. Ooops.
The code was then widely duplicated across multiple open source repositories, with one joker even opening up a pull request adding the source code to an Anthropic repo! Anthropic rapidly issues take-down demands to numerous repo owners, amusingly some described this as a vibe-takedown, due to messages being sent to repo owners who had not clnoed the code.
Wow.
Anyhow, back to the sourcecode. This blog post provides a detailed analysis of the more juicy parts, including, injecting fake tools to poison copycats, an undercover mode that hides the fact this is AI, meaning that AI-authored commits and PRs from Anthropic employees in open source projects will have no indication that an AI wrote them. There’s also a regex that detects user frustration, including terms like ‘piece of shit’.
Hilarious!
The Claude Code Leak
BUILD.MS
And while we were all rolling on the floor laughing at the horrors within this codebase, this post makes a very good point …
Yes, the code is a bit of a mess, but does that really matter? This post argus that it is not about the code, it is all about the Product Market Fit, and there is no denying hat Claud Code excels in this department.
Claw Code - a Rust re-implementation of Claude Code
GITHUB.COM
And the next turn in this saga, was the emergence of multiple rew-writes of Claude in various different languages, capitalising on the ability of AI (and Claude Code itself) to rapidly clone a project in a different language.
This one, a re-implementation in Rust, makes the claim that it is the fastest repo to reach 50k stars on GitHub, primarily demonstrating that stars are pretty meaningless. They were once considered an indicator of quality, but now are just a measure of the meme-worthy nature of a repo.
The rush of Claude Code rewrites once again highlights the controversial topic of “clean room”, and whether an AI model can be considered ‘clean’, and these rewrites considered legal (in that that they don’t violate copyright). A topic we touched on back in issue 34.
Interesting times!