sungjunlee

Improve an existing prompt or skill in place. Use for refine, sharpen, tighten, fix, upgrade, "feels off", or inconsistent artifact requests.

Create or refine spec/capabilities.md from repo signals. Use after spec-charter or for capability specs, component contracts, middle-layer specs, repo boundaries, 능력 명세, or grill.

Optimize prompts or skills with measured eval loops. Use for repeatable test inputs, eval runners, scoring, bounded mutations, or written stop conditions.

Decide where agent guidance belongs in a repo and propose the smallest harness change. Use for AGENTS.md/CLAUDE.md, hooks, subagents, MCP, plugins, sync, or pruning.

Create or amend spec/system-map.md. Use for architecture scope, system shape, runtime boundaries, core flows, invariants, storage/external systems, or SYSTEM_MAP.

Create, amend, or reassess spec/charter.md. Use for project direction, Objectives, Non-Goals, Decisions, stale spec findings, 기준, 헌장, 방향성, or spec axis.

Critique prompts or skills without editing. Use for review-only diagnosis, audit, "feels off" artifacts, or findings before a rewrite.

Design a skill-like artifact before writing it. Use for skill specs, suites, subagents, plugins, artifact-class decisions, triggers, file shape, or "meta-skill" requests.

Dispatch implementation tasks via worktree isolation. Use when delegating work to an executor, running background dispatches, or parallelizing independent tasks.

Use when completed relay executor work needs independent review against frozen Done Criteria — internal pre-publication rounds or post-publication PR rounds.

Drive already-planned relay leaves through fan-out, review, serial merge, crash recovery, and fleet status.

Interactive setup for relay routes. Use when the user asks to set up relay, configure company/personal relay routing, enable OpenCode or Pi, resolve model names, add provider/model routes, check advisory-review routing, or run relay-config doctor/check.

Use when a GitHub issue, sprint item, task description, or natural-language handoff should be implemented through autonomous executor dispatch; stops at ready_to_merge — merge only on explicit request.

Merge a reviewed PR, clean up worktree/branch, and close GitHub issues. Use after relay-review returns LGTM.

Use when a relay run needs its review anchor, scored rubric, or dispatch prompt — always before relay-dispatch, whether AC are explicit, vague, or missing.

Verify a task is ready to relay — score readiness on clarity, granularity, and verifiability, then ask bounded questions when the task is too ambiguous to plan.

Execute settled work in the current checkout with the lightest safe engine, escalating durable or high-risk work to relay before mutation.

Turn a vague idea into one selected direction and a compact Design Handoff, then recommend the next owner.

Compile settled intent into a tracker-neutral, implementation-ready Feature Spec with exactly one closing handoff.

Run a prompt on a chosen agent/model in the current directory and return its output with bounded non-interactive dispatch guardrails. Use for one-shot delegation through an installed provider CLI such as opencode, claude, codex, pi, reasonix, agent, or cline.

Review the current artifact with a real 4-6 person expert subagent panel. Use on explicit /gosu-review calls when code, plans, skills, design docs, decisions, or repo state need multiple context-specific perspectives.

한국 법령·판례·행정규칙·계약·컴플라이언스 질문에 사용한다. 사용자가 조문, 시행일, 판례, 과징금, 인허가, 노동·해고, 개인정보, 주총·등기, 중대재해, 계약서 위험 조항, 법령 변경, 한국 법률상 의무나 리스크를 묻는 경우 이 스킬을 사용한다. 공식 1차 소스 확인, 출처 권위 라벨, 자가 검증이 필요한 한국법 답변에 적합하며, 한국법 질문은 기억만으로 답하지 않는다.