@2008muyu/pi-plan

This Skill addresses the issue of excessive token consumption in AI coding agents when handling complex tasks. In real-world development, developers often need AI agents to perform deep analysis and planning. However, using a strong model (like Claude Opus) for the entire process can quickly escalate costs, while using a lightweight model directly for execution may lead to inadequate planning and lower code quality. By separating the process into two stages—first using a strong model for comprehensive codebase analysis and task breakdown, then employing a lightweight model for efficient implementation—it significantly reduces token usage while maintaining reasoning quality.

Usage is straightforward. Simply install this plugin via pi install and restart pi. You can then enter planning mode using the /plan command. The AI will automatically switch to your configured strong model, analyze the codebase in read-only mode, and call submit_plan to generate a plan file. After planning, you'll see a menu allowing you to either "Execute Plan" (switching to the lightweight model and restoring full tool permissions) or "Continue Refining" to further optimize the plan. Additionally, it supports /plan resume to restore a saved plan, /plan list to view all plans, /plan-settings for interactive configuration of model parameters, and the shortcut Ctrl+Alt+P to quickly toggle planning mode.

It is particularly suitable for teams or individuals who frequently use AI coding agents for complex development, especially those already using the pi coding agent and looking to control API costs without compromising code quality. For large projects, the plan-then-execute workflow can significantly boost efficiency and also facilitates team retrospectives and knowledge retention. Whether in full-stack, backend, or agent development, users can benefit from this approach.

It is recommended to configure the planning model (e.g., claude-opus-4-6 or deepseek-r1) and execution model (e.g., gpt-5.5 or claude-sonnet-4) uniformly via /plan-settings at the project's outset, adjusting the thinking level based on project complexity. Note that it primarily relies on the pi coding agent environment and is not compatible with other AI coding tools. During the planning phase, only read-only tools (read/grep/find/ls) are available; modification operations cannot be performed. Plan files are saved in the .plans/ directory and can be used for team collaboration or future review.