skills

Stress-test fuzzy intent into a code-aware, worker-ready plan or durable task.md Seed before execution. Use when scope, acceptance criteria, failure modes, data shapes, defaults, edge cases, or handoff boundaries are unclear, or when a durable Seed is requested. May precede decompose-and-dispatch when delegated or coordinated work is needed. Skip concrete plans with testable acceptance criteria and clear task boundaries.

Decompose a complex goal into independently executable work units and resolve each unit to an executor. Use when the user requests delegated, parallel, or multi-agent work, or when multiple units require dependency, ownership, or capability coordination. Skip ordinary sequential implementation by one agent.

Design or document non-trivial control flow with pseudocode and optional Mermaid diagrams. Use when the user asks for pseudocode or flow diagrams, when tracing an existing flow, or before implementing logic with interacting branches, multiple side effects, lifecycle, ordering, or concurrency risk. Skip local guard clauses, one-line edits, config-only changes, and established local patterns.

Review a code change for unnecessary abstractions, duplicated durable policy, compatibility cascades, oversized integration-test matrices, and scope expansion. Use during implementation when one new abstraction causes three or more follow-up regressions, or after tests are green and before commit when a change adds persisted fields, interface methods, lifecycle states, compatibility branches, destructive-operation safeguards, or substantially more test code than production code; keep the review read-only unless simplification is explicitly requested.

Capture reusable learnings from completed and verified work — tasks, reviews, bug fixes, architecture decisions, debugging sessions — into task-scoped notes and durable solution docs so future agents skip re-discovery. Use for "compound learning", "document what worked", "capture a reusable pattern", "summarize lessons", "update solution docs", "close the loop". Do not use for in-progress work, uncertain outcomes, trivial fixes, performing code review itself, or replaying a session as an executable recipe (that is `session-recipe`).

Execute exactly one assigned dispatch packet or atomic work unit, staying inside allowed scope, respecting non-goals and forbidden scope, verifying the result, and returning a structured completion report. Use when an agent, worker, or subagent is handed a bounded task with fields such as unit_id, objective, allowed_scope, dependencies, acceptance_criteria, verification, or return_contract — typically a packet produced by decompose-and-dispatch.

Prepare code-review context briefs, PR/diff context packages, severity rubrics, false-positive suppressions, reusable reviewer instructions, and AI reviewer prompts. Use for AI review preparation, PR or diff review context, review invariants/non-goals, P0/P1/P2/P3 calibration, and instructions for Codex, GitHub Copilot, CodeRabbit, Cursor, or other reviewers. Do not use when the user asks to perform the code review itself.

Set up and verify ground-truth session recording, distill a finished session into a replayable recipe of dispatch packets, and replay a saved recipe step by step. Use when the user asks to record a session, turn completed work into a recipe, replay a recipe.yaml, validate a recipe file, or set up session-recording hooks.

Study an open-source codebase using a four-mode (Diátaxis-based) question structure that keeps each question to a single learning intent. Use when the user wants to study an unfamiliar repo or library for learning (e.g. "이 오픈소스로 스터디하자", "이 라이브러리 어떻게 동작하는지 공부하자"), or to understand a dependency's internals before adopting it. Not for shipping a fix.

Coding quality guardrails for AI agents working on Go, Python, Java/Kotlin, or TypeScript. Use when Codex is coding, reviewing, refactoring, fixing tests, debugging regressions, preparing AI coding-agent prompts, or evaluating changes where AI slop, test gaming, over-abstraction, weak verification claims, or quality debt must be prevented.

Shared discipline for designing deep modules, small interfaces, clean seams, adapters, test surfaces, and architecture improvements. Use when designing or reviewing module boundaries, refactors, controller/service composition, interface shape, testability, coupling, dependency injection, or when another skill needs architecture vocabulary for deepening code.

Disciplined diagnosis loop for hard bugs, regressions, flaky behavior, failing tests, broken UI flows, performance regressions, and user-reported failures. Use when the user asks to debug, diagnose, investigate, reproduce, find root cause, or fix something broken, slow, failing, throwing, intermittent, or behaviorally wrong.

Build and sharpen a project's domain language and decision record. Use when terminology is ambiguous, the user wants canonical terms, glossary work, DDD-style ubiquitous language, ADRs, context docs, naming alignment, or when planning/design work depends on precise domain concepts before implementation.

Review, create, or refine portable agent skills for predictable behavior, progressive disclosure, clean invocation metadata, concise instructions, reference splitting, trigger wording, and no-op pruning. Use when writing SKILL.md files, porting skills across Codex, Claude, Gemini, or other agent runtimes, auditing skill quality, or improving skill context load and execution reliability.