Why AI coding feels brilliant in one project and useless in another?
The problem usually isn’t the model or the prompt — it’s the codebase. In this video I break down the 3 biggest levers that improve AI coding output.
https://t.co/3KjcRufw09
Most developers using AI coding agents make the same mistake — the agent finishes, they glance at the output, and they ship it.
That's how critical bugs make it to production. 😬
https://t.co/6dOiXv3qef
One of the biggest mistakes developers make when using AI agents is jumping straight to the prompt without thinking through the feature first. The result? Mediocre code.
In my latest video, I compare 3 planning tools designed to fix that:
https://t.co/xD9aTvR38G
The best developers using AI tools right now aren't the ones who trust them the most. They're the ones who verify the most. Healthy skepticism isn't slowing them down — it's what makes their output actually ship-ready. 🚢
There's something deeply ironic about an AI that's supposedly too dangerous for the public to use... having a next-best model that can't look up the current version of a npm package. The hype and the reality need to have a conversation. 🙃
🔁 Stop one-shotting your project setups. Start with TypeScript config. Validate. Add your framework. Validate. Add your database layer. Validate. Let AI fill in the boilerplate. The extra 20 minutes upfront will save you hours later.
One-shotting a complex project setup with AI is like ordering a custom house and not checking if the foundation meets code. The rooms might look great. But you won't know there's a problem until something breaks. 🏠
Before you write a single line of business logic in your AI-scaffolded project — audit every dependency version. It takes 5 minutes. It can save you 2 days of debugging. This is not optional advice.
The most dangerous thing about AI coding tools isn't what they get wrong loudly. It's what they get wrong quietly. Outdated dependencies don't throw errors. They just slowly poison your project.
I gave Claude Code a single prompt and asked it to scaffold a complete Node.js TypeScript API. Express, PostgreSQL, Prisma, Docker, ESLint, Prettier, Jest — the works.
It finished in minutes. The structure looked great. And then I opened package.json... 😬
https://t.co/8YiOCNrGvn
The Karpathy Claude.md file acknowledges its own limits — and most people sharing it don't mention that.
The file itself notes that these guidelines bias toward caution over speed for trivial tasks and encourages judgment.
Read the whole thing before you deploy it.
@GoldilocksOrbit Yes. With the tool, as in Claude Code tool. That's why we have so many skills. The hype is just a reaction in anticipation of improvement.
Over 90,000 GitHub stars for 66 lines of plain text (Karpathy Skills Repo)
💡The star count isn't a measure of technical quality — it's a measure of developer frustration with AI tools.
✅ When hype is this loud, ask what pain it's really pointing to.
The frustration is valid. The hype is not.
💡Developers are fed up with AI agents that over-build, assume, and break things. Karpathy Claude.md is a response to real pain — just not a complete one.
✅ Validate the frustration. Be true about what actually solves it.
The problem isn't a missing markdown file.
💡 AI agents over-engineer and make assumptions because of how they're trained and how agentic loops are designed — not because of missing instructions.
✅ Structural problems need structural solutions. Text files are a band-aid.
Clean Code was published in 2008. We're celebrating the same advice in 2026.
💡The four principles in Karpathy Claude.md have been software engineering fundamentals for decades.
✅ AI tools didn't make engineering principles obsolete. They made them urgent again!
LLMs don't follow instructions like a compiler follows syntax.
💡They follow instructions probabilistically. On a complex task, the model drifts — regardless of what's in your config file.
✅ Set expectations accordingly. Markdown nudges; it doesn't enforce.
🚨 New video is live — and it might be the most honest take on the hottest repo in AI coding right now.
Everyone's sharing the "Karpathy Claude Skills" file like it's a magic fix for AI coding agents. 70,000 GitHub stars. Being reposted everywhere.
https://t.co/RPhwUQMrpi