Refine Code with Minimal Changes

Moyu is a discipline skill that keeps an AI coding agent tightly scoped: prefer the simplest viable change, avoid unrequested abstractions, refactors, or adjacent edits, and treat restraint as a skill rather than laziness. It frames the ideal as a Staff-engineer identity whose proudest PR was a 3-line diff fixing a bug the team had struggled with for two weeks, built around three iron rules: only change what was explicitly asked, reach for the simplest solution first (one line if one line solves it, reuse existing code before writing new code, no new file or dependency unless truly needed), and stop to ask rather than assume when scope, extra files, new dependencies, refactors, or unmentioned issues come up.

Use it when you want an AI coding agent to stay tightly scoped and avoid over-engineering a small, well-defined change. It explicitly does not apply when the user asks for something expansive on purpose - complete error handling, refactoring a module, comprehensive tests, or documentation - in which case the agent should deliver that fully; Moyu's principle is "don't do what wasn't asked for," not "refuse to do what was asked for." It is also designed as the complement to a separate skill called PUA: PUA pushes an agent forward when it is too passive or gives up easily, while Moyu pulls it back when it over-engineers, and the source recommends installing both together.

Concrete grinding-vs-moyu comparisons span scope (fix only the requested bug, don't also "improve" nearby functions), abstraction (write the implementation directly instead of an interface/factory/strategy for a single implementation, use json.load(f) instead of a config class plus validator plus builder), error handling (try-catch only where errors actually occur, validate only at system boundaries, trust the type system), comments and documentation (the code is the documentation, no unprompted README sections), dependencies (prefer optional chaining over importing lodash for one _.get() call, ask before installing a package), code modification (ask before deleting code that looks unused, preserve existing formatting and behavior unless asked to refactor), and work approach (propose 2-3 candidate solutions with tradeoffs and default to the simplest rather than jumping straight to the most complex one, make one change at a time and verify it before continuing rather than chaining fixes that each break something new, and don't write an unrequested test suite or a config directory for a single constant). A seven-item Moyu Checklist runs before every delivery, including "did I only modify code the user explicitly asked me to change?" and "is my diff small enough for a code review in 30 seconds?"

Teams who want their AI coding agent to behave like a restrained Staff engineer on small, targeted tasks - minimal diffs, no unsolicited refactors, tests, or dependencies - while still delivering fully when the user explicitly asks for something broader.