Add structured planning to your Cursor or Windsurf editor so it writes a plan before touching any code
Let your AI editor browse web pages and run DuckDuckGo searches autonomously during a coding task
Set up a two-model multi-agent workflow where one AI plans and another executes to improve output quality
Configure your AI editor to accumulate project-specific lessons so it repeats fewer mistakes over time
This repository is a configuration kit that adds more capable AI behavior to Cursor, Windsurf, and GitHub Copilot, the AI-assisted coding editors. The goal is to get those tools to behave more like Devin, an AI coding agent that was notable for writing out a plan before starting work, updating that plan as it went, and learning from corrections along the way. Devin charges per task, whereas Cursor costs a flat monthly fee, so this project aims to unlock similar behavior at a lower price. The main file you copy into your project is a rules file (.cursorrules for Cursor users.windsurfrules for Windsurf users, or a Copilot instructions file for GitHub users). These files tell the AI editor how to behave: write a plan first, revise the plan as things change, and record lessons learned whenever you correct a mistake. That last part is what the README calls self-evolution: the AI accumulates project-specific notes over time so it makes fewer repeated errors. The repository also includes a small set of Python tools that extend what the AI can do during a task. These tools let it browse web pages using a browser automation library called Playwright, run search engine queries through DuckDuckGo, and call other AI models to analyze text or images. The AI decides on its own when to use these tools based on what the task requires. There is also an experimental multi-agent mode on a separate branch. In that setup, one AI model acts as a high-level planner and another handles the step-by-step execution. The idea is that having two separate models check each other improves the quality of the result. Setup takes about a minute. The recommended path uses a tool called Cookiecutter that prompts you for a few choices and copies the right files into a new project folder. The project is released under the MIT license.
← grapeot on gitmyhub — every repo by this author, as a profile.
Verify against the repo before relying on details.