Developers can guide AI coding assistants through structured workflows for complex tasks
If you want AI to follow verified steps rather than giving vague answers when fixing code or finding bugs, this repository is suitable. It collects work instructions and automation rules that apply to multiple AI coding tools. You state the goal, and the AI selects the necessary tools, executes them, and verifies the results on its own.
It handles things automatically when you state the goal
The core of this repository is the poteto-mode command. When a user describes a problem or goal instead of a specific solution, this command finds and executes the appropriate workflow. For example, if you ask it to fix a bug, it first reproduces the issue, analyzes the cause, delegates the fix, and finally checks if the previously failing tests pass. If the fix involves significant changes crossing function boundaries, it automatically inserts an architecture review step before implementation. This increases the likelihood that the AI provides verified results rather than just changing code.
Verifying problems that are hard to catch with tests
When dealing with code that is difficult to verify with standard tests, such as concurrency issues or complex state changes, you can use the formal verification feature. This feature selects the most dangerous properties in the code and verifies them using specialized tools like TLA+ model checking or Lean proofs. After verification, the method is saved to the project so it can be reused when handling similar problems in the future. This helps the AI go beyond simply writing code to ensuring logical correctness.
Works with various AI coding tools
This repository is not tied to a specific AI tool and is designed to work in multiple environments, including Claude Code, Codex, GitHub Copilot, and Gemini CLI. It converts commands and settings to match the characteristics of each tool, so the installation method differs depending on the tool you use, but the user experience remains similar. Some tools automatically apply workflows after installation, while others may require users to enter commands manually. Since not all tools operate in the same way, it is important to check the installation instructions for your specific tool version.
Where the data goes
This repository has no server or telemetry features. All data read by the AI, including session logs, is sent only to the model provider you select. Scripts run locally, and GitHub-related tools use your login credentials. If you use the show-me-your-work feature to record the work process, you can track the decision-making process, but these records are not committed by default. If you upload these records to a shared PR, anyone with access to the PR can see them, so be careful.
Things to check before use
This repository is released under the MIT license, so you can use it freely. However, compatibility may vary depending on the version of each AI coding tool, so you must check the minimum version requirements specified in the README. Specifically, tools like Copilot or OMP are noted to operate stably only within certain version ranges. Additionally, the automatic routing feature does not apply to all tools, so you should check if your tool supports automation; if not, keep in mind that you will need to enter commands manually.