vikingmute/shifu

Shifu is an agent skill for coding assistants that splits a big task into small, self-contained written specs so that cheaper AI models can carry out the actual work. The idea behind it is to save your most capable, most expensive model for the part that benefits most from intelligence: understanding the whole task, deciding how to break it apart, and writing instructions that leave no room for confusion. Once those specs exist, a cheaper model does the implementing, testing, and shipping. Shifu itself never writes application code. The specification document it produces is the actual output. You install it with a command line tool called skills, and it works in any coding agent that supports the Agent Skills format, such as Cursor, Codex, or Claude Code. You interact with it through commands like asking it to analyze and plan a task, asking it to execute a specific plan with a chosen model, asking it to review an existing spec, or asking it to check the status of everything in progress. Shifu works in two ways. In agents that can spawn their own subagents, it can run the whole plan, dispatch, and review cycle automatically without you leaving your session, picking cheap workers for each piece and reporting a verdict on each one when they finish. In agents without that support, it writes plain markdown spec files into a plans folder that you can hand to a separate, cheaper session yourself. When breaking a task apart, each sub-task is assigned one of three strategies: a direct approach for straightforward sequential changes, an isolated git worktree for higher risk work like large refactors, or a read only exploration mode for research and analysis. Every spec written this way includes all the necessary context inline, a list of what is explicitly in and out of scope, and verification steps with expected results so that completion can be checked mechanically rather than by judgment call. Shifu enforces a few hard rules on itself: it never edits your source code directly, it never runs commands that change your working tree, and it never repeats secret values it comes across, only their type and location. Shifu is released under the MIT License.