gprecious

Convert raw feature requirements, bug reports, or rough task descriptions into a goal-optimized prompt for the agentic coding tool that will run it — Claude Code or OpenAI Codex. 거친 요구사항·버그 리포트·대충 적은 작업 설명을, 그 작업을 실제로 실행할 코딩 에이전트(Claude Code 또는 Codex)에 맞춘 goal 최적화 프롬프트로 변환한다. Use when the user wants a requirement shaped into a prompt/spec/goal for an agent — "이거 프롬프트로 만들어줘 / 변환해줘", "코덱스(codex)/클로드코드용으로", "목표·범위·완료조건 갖춘 프롬프트로" — or mentions the goal feature or /goal command. By default targets the agent running the skill (Claude Code → Claude-optimized, Codex → Codex-optimized); honors an explicit named target. Proactively fire on vague requirements headed to a coding agent ("대시보드 빠르게 해줘", "add auth", "로그인 버그 고쳐줘"). Not for implementing/debugging it yourself, codex code review, explaining /goal, translating, editing AGENTS.md, multi-agent orchestration (cmux/herdr), or writing a plan.

Use when the user wants to turn a repetitive, reviewable task into a controlled agent loop — phrases like "이거 자동화/반복하고 싶어", "루프로 돌리고 싶어", "정기적으로 점검하게 해줘", "loop engineering", "agent loop 설계", or wants to audit existing loops ("루프 점검", "loop audit"). Designs a loop with trigger·state·verification·human-gate·stop + 7 guardrails + 6 metrics, maps it onto an existing engine (herdr/loop/Cron/Workflow/Codex Automations), and writes a spec to docs/loops/. Does NOT execute loops itself and is NOT for one-off prompts (use goalcraft) or multi-agent run orchestration (use herdr/cmux).

Convert a recurring or repetitive task description into a Claude Code /loop prompt — the loop sibling of goalcraft. 반복·주기적 작업 설명을 Claude Code /loop 용 프롬프트로 변환한다(goalcraft의 loop 형제). Use when the user wants a task shaped into a /loop prompt — "루프 프롬프트로 만들어줘", "/loop 용으로", "주기적으로 반복하게", "매 N분마다 ~", "~까지 반복해서", "이거 loop으로 돌려줘" — or mentions the /loop command, recurring/scheduled in-session execution, or babysitting/polling a long-running thing. Picks the mode (fixed interval → cron, self-paced → model self-paces, or loop.md default) and bakes in an objective stop condition, an iteration ceiling, and a verification check. Proactively fire on recurring-execution requests ("배포 끝날 때까지 확인", "CI 통과할 때까지 반복", "keep checking the deploy"). Not for single-shot tasks or /goal until-condition prompts (→ goalcraft, which owns /goal), running the loop yourself, deciding whether something should be a loop or auditing loop governance (→ loop-engineering), translating, or explaining /loop.

Use when creating, revising, or critiquing frontend UI (web or mobile) where visual quality, product fit, design-system fidelity, or avoiding generic "AI-slop" output matters — e.g. "make this look good", "why does this look AI-generated", building a landing page / app screen / dashboard, or a design pass on existing UI. Turns vague design intent into a decision-first workflow: classify the surface, decide direction before code, escape default looks, build to measurable craft standards, and verify the rendered result. Portable across Claude Code and Codex.

Use when the user asks to inspect, summarize, explain, configure, connect, curate, or troubleshoot the session-journal Obsidian logging workflow for Claude Code or Codex — including pointing it at an existing/Obsidian-Sync vault, keeping AI-generated notes distinct from hand-authored ones, multi-machine setup, distilling durable knowledge, or verifying the vault config.

Use when the user asks to save durable lessons from session-journal into the LLM-Wiki, collect mistakes or insights from agent work, create reviewable wiki draft candidates, or report those draft candidates to Slack.

Use when the user wants to add or generate an E2E / verification test harness for a project — especially "에이전트가 작성·검증·자가수정하는" Playwright E2E for web apps (Next.js/Vite/Remix/SvelteKit), or a verification-harness methodology for non-web (CLI/API). Detects the stack, picks a recipe, co-authors critical flows with the user, scaffolds config+seed+specs+heal docs, wires an AGENTS.md pointer so future agents auto-run the harness after changes, and runs it. Works in Claude Code and Codex.

Use when the user explicitly asks for herdr-based multi-agent orchestration, multiple visible Claude/Codex worker panes, or cross-CLI comparison inside herdr, especially when workers need distinct CLI/account/browser state.

GStack + GSD + Superpowers + build-loop을 cmux pane + codex CLI 기반으로 통합하는 자율 빌드 파이프라인. /build <topic> 한 줄로 spec → 페이즈 분해 → TDD 루프 → 통합까지 진행. claude orch가 plan/refactor를 담당하고 long-lived codex worker가 코드 작성. 컨텍스트는 /compact로 관리.

Iterative codebase improvement orchestrator. Audits → prioritizes → auto-fixes Priority 1-2 via category-split PRs → tracks plateau across iterations. Independent of harness-orchestrator.

피처 개발 하네스 오케스트레이터. /feature 명령으로 TDD 기반 Plan-Contract-Test-Build-Evaluate-Integrate-Learn 사이클을 실행. /feature --init, /feature --resume, /feature "설명" 을 요청할 때 사용.

Use when an agent needs background or prior findings on a topic that may already be researched — look it up in the LLM-Wiki (the Obsidian vault built by research-engine) instead of re-researching. Searches distilled concepts/entities and verbatim reports, returns a cited answer, never edits the vault. Triggers on "위키 찾아봐", "이전 리서치 있어?", "wiki lookup", "what do we already know about X", needing context already captured by /research.

Use when the user asks Codex to research a URL, paper, repo, video, blog, documentation page, or topic and produce a structured cited markdown report or follow up on a previous research session using the research-engine workflow.

ulw 또는 ultrawork 키워드로 활성화되는 병렬 에이전트 오케스트레이션 모드. "ulw:", "ultrawork:", "ulw", "울트라워크" 입력 시 사용. 모든 외부 도구 총동원, 완료 강제 모드. oh-my-opencode의 Ultrawork/Sisyphus 패턴 구현.