your claude.md is probably doing more harm than good

I've been on a hunt for the best claude.md structure

Howdy, ya'll -

I've been on a hunt lately to find the best structure for a Claude markdown file

It's part of the upcoming relaunch of my paid "production vibe coding" community (shameless plug, link at way bottom of email)

After a lot of research, wasted tokens & testing, here's what I'm landing on:

First things first, around 300 lines max. If for some reason it needs to be larger than that, you should create a separate "rules" directory and tell Claude in your claude.md file when to read it - basically a "rule router".

Here's the ideal structure:

• One-liner with project motivation/direction
• Non-obvious tooling, settings, project/version specific quirks
• A CONCISE architectural map (things that aren't obvious / would take reading multiple files to discover)
• Rules with verifiable instructions (e.g., parameterize all SQL, not "write clean safe code")
• Hard constraints / anti-patterns
• Pointers to deeper docs (shout out to Intent Layers)
• Gotchas / Tribal Knowledge (e.g., "retry logic looks broken, but it's load bearing. Check PR #123 for context")

Point being - a lot of the old conventions and ways we used to stuff Claude.md files aren't really that effective anymore.

NEWS WORTHY

Karpathy joining Anthropic

https://x.com/karpathy/status/2056753169888334312

Another founding team member from OpenAI has now left for Anthropic.

And this one is a doozy!

Andrej Karpathy has joined the R&D team at Anthropic, focusing on the pre-training work.

Pre-training is how models get their core knowledge and capabilities.

Given how market-shifting his opinions and approaches to things tend to be, I think this signals sunny horizons for Anthropic.

Figma Integrates Agent directly

https://www.figma.com/blog/the-figma-agent-is-here/

One of the dominant patterns from the last few months with vibe-designing has been giving agents direct control over the design canvas.

The general principle: "anything a human could do, an agent can do"

This is why tools like Pencil.dev and Paper.design have stolen the hearts of vibe coders everywhere.

Well, Figma just released an agent-native canvas functionality where you can do all of that stuff in Figma directly.

Some highlights:

• Exploring directions - you can ask it to explore fundamentally different approaches to your features/designs
• Remixing styles & screens - it can take existing stuff and give it some flair
• Bulk updates - if you want to apply a fix across all of your designs, you can do it pretty simply

It works on a credit-based system, so it's worth seeing where the market takes it!

Mythos Getting Tested In The Wild

https://blog.cloudflare.com/cyber-frontier-models/

It's been a few weeks since we've heard anything meaningful about the new Mythos model from Anthropic, which was deemed "too dangerous" to be used by us normies.

Well a few companies (Cloudflare included) have been testing it out, and the claims seem to be backed by real-world use cases.

Cloudflare claims it's exceptionally good at finding security vulnerabilities, particularly those caused by chains of failures.

A lot of attacks that sophisticated bad-guys use often combine exploits together, and Mythos is apparently really good at finding those chained-attacks.

If you want a really light-weight version of finding that type of stuff in your own apps, I'd recommend you check out Deepsec by Vercel (link below).

I did a video on it here:

WORTH CHECKING OUT

CodeGraph: something that's starting to "go viral" is the idea of creating knowledge graphs for your codebase. Meaning, you use tools to build a deep contextual understanding of how all the pieces connect, and then your LLM uses that map to navigate better. The results are pretty solid, particularly around SPEED. Meaning, it tends to get to solutions much faster. CodeGraph is one such tool, you should check it out

Academic Research Skills: when I'm building projects in domains I don't really know about, exploration & context gathering is really important. This skill helps with that by codifying an "academic research" pipeline that can surface grounded info on best practices and considerations. Super valuable if you're someone that likes to push the envelope with what you build. Check it out here

Trail of Bits: as part of my paid community re-launch, I've been compiling a lot of open source skills into master-plugins for specific tasks. One of those tasks is security auditing your app and your new features. Trail of Bits is one of the libraries I use a lot, it's worth checking out if you... ya know, care about the security of your app and customers.

DeepSec: I talk about this tool a lot because I'm still shocked at how something so good is open source. It's called DeepSec, and it's a deep-security harness built by Vercel. It uses your Codex/Claude CLI (or their API if you want) to scan your project for really hard to catch security issues.

Happy Vibin'

-Sean

PS - major relaunch of my paid "production vibe coding" community is starting to drop today. Price is going to increase in the next 2-3 weeks, so if you've been on the fence I'd join today: https://www.skool.com/tech-snack-pro