hei5enbug-agent-setup

Plan a large effort that spans more than one agent session as a shared map of decision tickets, then resolve the tickets one at a time until the route to the destination is clear. Use only when the user explicitly asks to create or continue a decision-navigator map.

Conduct a rigorous Socratic requirements interview before implementation. Cover topology discovery, evidence-backed questions, ambiguity scoring, contradiction handling, closure gates, and an execution-ready specification. Use only when the user explicitly asks to run a deep interview, by naming the skill or by invoking its command. Do not start it because a request looks vague, ambitious, risky, or underspecified, and do not offer to start it unprompted. Works across agent hosts with structured-question, exploration, independent-worker, and plain-text fallbacks.

Rewrite existing text so it reads naturally while its meaning stays exactly the same. Use when a document, message, review comment, commit message, PR description, chat reply, or file is supplied and the request is to polish, tidy, smooth, restructure for readability, fix awkward wording, or remove AI-sounding and translated-sounding Korean, including "다듬어줘", "자연스럽게 고쳐줘", "읽기 쉽게", "가독성 개선", "문장 정리", "말이 이상해", "AI 티 제거", "사람이 쓴 것처럼 윤문", "번역투 제거", "ChatGPT 문체 수정", "한국어 휴머나이징", "장르·강도 조정", "polish this", "make it read better", "reword". Also use on an agent draft before it is posted or committed. Do not use to write new text, to translate, to summarize, to expand with added explanation, or to check whether the content is correct, and never promise that a detector will be defeated.

Convert Markdown, HTML, PDF, DOCX, or Google Docs content into a Confluence page and keep the page synchronized with later source revisions. Use when importing, publishing, migrating, updating, or repairing documents in Confluence, including tables, links, images, attachments, diagrams, page structure, or text fidelity.

Design, revise, validate, and export clear static flow diagrams across SVG, HTML/CSS, Figma, FigJam, draw.io, Mermaid, and raster deliverables. Use for flowcharts, process maps, architecture flows, decision flows, diagram-system unification, spacing and routing repair, overlap removal, accessible labeling, viewBox tightening, or PNG export correction. Works across agent hosts by separating the semantic graph from tool-specific rendering and using capability-based fallbacks.

Update hei5enbug-agent-setup in Claude Code and Codex, then apply the update to every idle Orca-managed session without restarting it: reload plugins in Claude Code sessions and compact the sessions whose loaded instructions are stale. Use only when the user explicitly asks to refresh or apply this plugin's update in running Orca sessions.

Create, modify, validate, and evaluate agent skills across AI agent hosts. Use when users want to build a skill from scratch, improve an existing skill, test skill behavior against a baseline, benchmark quality and variance, package a skill, or optimize its description for reliable triggering. Apply this skill to portable SKILL.md instruction bundles as well as host-specific skill packages, while preserving the target host's required metadata and conventions.

Quickly analyzes staged and unstaged changes together, or the scope the user names, plus recent commit history, then recommends 5 commit messages that match the repository's existing subject language and style and always open with the type prefix. When the user asks to create a commit, it instead commits that change set with the single most appropriate subject, without asking and without any trailer.

Create, improve, or review technical design documents, architecture proposals, RFCs, system designs, data designs, API designs, and design outlines. Use when the user explicitly requests design work or directly invokes this skill, with phrases such as "설계 문서 작성", "설계 문서 개선", "설계 문서 목차", "design doc", or "RFC review". Enforce evidence, complete design coverage, explicit trade-offs, precise contracts, predictable structure, natural Korean, effective tables and diagrams, and stable references. Do not use for translation, generic copyediting, implementation planning, user guides, release notes, general implementation, diagnosis, summarization, or code review without explicit design intent.

Runs a bounded, evidence-based debate between the current agent host and a dedicated Claude or Codex opponent session. The current host participates directly and keeps its selected model and reasoning effort; the opponent uses xhigh by default. Runs two exchanges by default and up to five when the user explicitly requests it. Shows concise progress, supports durable execution and reconnection, applies agreed improvements by default, and writes unresolved questions separately. Leaves every file unchanged when the user requests a synthesis only. Never trigger automatically. Use only when the user explicitly asks with phrases such as "두 에이전트로 토론시켜", "Codex와 토론", "반대 의견도 받아 겨뤄봐", "적대적으로 검토", or invokes /tiki-taka directly.