Vibecode Starter by Prototyper.co

You open Claude or Cursor, start a conversation, build half a feature, then close the session. Tomorrow you're back to square one. The assistant doesn't remember what it built. You paste in the same context. You explain the codebase again. Repeat. This repository helps you avoid that.

What's actually going wrong

AI assistants have no memory between sessions. They can't see files on your machine unless you show them. When you run multiple agents, they can't talk to each other. The technology is good. The workflow is broken.

Most tutorials tell you to "write better prompts." That's not the problem. The problem is organizational. You need a place to put context, a way to break work into pieces that survive session boundaries, and a channel for agents to coordinate.

Three folders

context/     Put reference files here before a session
plans/       Break projects into numbered prompts
messages/    Agents leave notes for each other here

That's the whole system.

context/ holds specs, docs, schemas, screenshots. Drop files in before you start. Agents can read them. Gitignored so your repo stays clean.

plans/ structures work that takes multiple sessions. A plan is a folder with a README explaining what you're building and a series of numbered prompt files. Agent A runs prompt 01, agent B picks up at prompt 02. Nobody loses track.

messages/ is a bulletin board. An agent hits a blocker, it writes a file. Another agent (or you) reads it and responds. When the issue is resolved, delete the file. Simple.

Getting started

git clone https://github.com/PrototyperAI/vibecode-starter
cd vibecode-starter

Then add your own code, drop relevant docs in context/, and start working. The AGENTS.md and CLAUDE.md files give AI assistants instructions. Read them if you want to customize behavior.

Why files

Because everything can read files. Claude can. GPT can. Cursor can. Your terminal can. Git can track them. You can edit them in any text editor. No plugins. No accounts. No API integrations.

Files also force you to be explicit. When context lives in a folder, you know exactly what the agent sees. When plans live in numbered prompts, you can review the approach before any code gets written. When messages live in markdown, you have a paper trail.

Who this is for

Developers or people building real products with AI assistance who got tired of the groundhog-day workflow. Works with any AI coding tool. Designed to get out of your way.

License

MIT