A place for developers to talk about building on Farcaster.

/dev

🚨 Breaking News for Developers 🚨

The other day, I faced a task: to develop a UI for a bridge from SOL to EVM with a non-trivial token retrieval process. Via protocol vibes... 

I turned to the brand-new CLI from Anthropic – Claude Code. But first, the key to success? Setting the right context & prompt for the LLM. Without it, you're just guessing. Trust me, architecture and constraints are a must. 🧠

dev but triyng to retire | always a gambler, sometimes a holder |

pfp: @doginaldogsx

In addition to the API spec, deBridge has over 20 HTML pages with best practices and full protocol details.  Moving all of that into the prompt context? Total headache. Losing formatting, copying content without references... it felt like "dancing with tambourines." 

And then, a lightbulb moment 💡: Modern documentation is NOT adapted for LLMs. It’s time for change.

Whether you like AI or not, LLMs can drastically simplify the lives of developers. The best part? You don’t need to be a genius coder – just vibe code it! 🤖

A few days later, Andrey Karpaty (the vibe-coding king himself) posted about the same thing on X. We’re on the same wavelength! 🔮

Some of the projects already embracing llm.txt: • Svelte: The ultimate frontend framework: Svelte llms.txt
• Anthropic: For Model Context Protocol: Model Context Protocol
• viem: My crypto favorites: Viem llms.txt

Bonus: Viem's own framework, vocs, auto-generates LLM files for docs. Pretty epic. 🔥

There’s a solid guide for developers on how to prepare documentation for LLMs: llmstxt.org 

But here’s the kicker: Current documentation systems still suck. 

Sure, you can add a URL in Cursor, but it only parses a single URL. Subpages? Not so much. But with llm.txt, Cursor can retain ALL the context, create embeddings, and refer to the docs seamlessly. No more nonexistent methods or random parameters. 🙌

Many vibe-coders think LLMs will "figure it out" on their own... but NO! 🎯 You need to guide it, otherwise it’ll wander aimlessly. I can share my findings after 1.5 years of full-time LLM dev if you're interested. Just ask!

The real MVP of my journey was the JSON with the OpenAPI spec of the deBridge API. It was a game changer. With one prompt, Claude Code generated the integration class and typed all the necessary parameters. 

But then… problems started. 😬