Polish AI-written Korean documents into natural human-sounding Korean by editing files in place. Removes translation-ese (์ง์ญ์ฒด) first โ catalog patterns get phrase-level substitution, principle-based judgment handles sentence-level rewrites โ while never touching code, identifiers, quotes, or terminology. TRIGGER when: user asks to polish, naturalize, or fix awkward Korean in a document file (e.g., "์ด ๋ฌธ์ ๋ค๋ฌ์ด์ค", "๋ฒ์ญํฌ ๊ณ ์ณ์ค", "ํ๊ตญ์ด ์์ฐ์ค๋ฝ๊ฒ ๋ฐ๊ฟ์ค", "AI ๋งํฌ ์ข ๊ณ ์ณ์ค", "๋ฌธ์ ๊ต์ ํด์ค", "์ด ๋งํฌ๋ค์ด ํ๊ตญ์ด ์ข ์ฌ๋๋ต๊ฒ", "polish this Korean doc", "fix the translation-ese"). Also trigger with /doctor-korean-document slash command. DO NOT TRIGGER when: user wants content changes, summarization, or translation to another language; the text is not in a file (inline/pasted text is out of scope); the document is not Korean; or the user only wants spelling/spacing lint (use a dedicated linter).
Auto-create or update GitHub Pull Requests with conventional commit-style titles. Detects base branch, generates PR title and description from branch commits, links related issues, and creates the PR via GitHub API. TRIGGER when: user asks to create a PR, open a pull request, push changes for review, or send work upstream (e.g., "PR ๋ง๋ค์ด์ค", "PR ๋ ๋ ค", "ํ๋ฆฌํ ์์ฑํด์ค", "ํ๋ฆฌํ์คํธ ์ฌ๋ ค์ค", "๋ฆฌ๋ทฐ ๋ณด๋ด์ค", "๋ฆฌ๋ทฐ ์ฌ๋ ค์ค", "์ด๊ฑฐ PR ํด์ค", "ํธ์ํ๊ณ PR ๋ง๋ค์ด", "์ฝ๋ ๋ฆฌ๋ทฐ ์์ฒญํด์ค", "์ด ๋ธ๋์น ์ฌ๋ ค์ค", "create a PR", "open a pull request", "send this for review", "push this up for review"). Also trigger with /github-pr slash command. DO NOT TRIGGER when: user is reviewing an existing PR, asking about PR status, or discussing PR workflow without intent to create one now.
End-to-end 5-phase pipeline from idea to code-ready state โ analyst intake โ ralplan โ ralph/team โ verifier-gated QA โ multi-perspective validation. Artifacts land in .dt-handoff/<slug>/. Auto commit/PR PROHIBITED. TRIGGER on "/autopilot", "autopilot", "build me", "create me", "make me", "I want a/an". DO NOT TRIGGER for brainstorming, single bug fixes, or quick changes.
Interactive requirements-clarification interview that turns a vague request into a pipeline-ready spec. Delegates to the `analyst` agent to surface ambiguities, hidden assumptions, and gaps, then resolves each with the user via `AskUserQuestion` โ offering concrete candidate answers (with a marked default where one clearly fits) and batching independent clarifications into one call โ and writes a clarified `spec.md` that `/autopilot` and `/ralplan` can consume. TRIGGER when the user has a fuzzy idea/spec/request and wants the ambiguity resolved into clear requirements โ "clarify this", "ask me what you need to build this", "interview me to nail down requirements", "์ด๊ฑฐ ์ ๋งคํ๋ฐ ์ง๋ฌธํด์ ์ ๋ฆฌํด์ค", "์๊ตฌ์ฌํญ ๋ช ํํ ํด์ค", "๋ชจํธํ ๋ถ๋ถ ์ง์ด์ ์คํ์ผ๋ก ๋ง๋ค์ด์ค", "/interview". DO NOT TRIGGER when the request is already unambiguous (go straight to /ralplan or /autopilot), when the user wants to reach their OWN design decisions by being questioned without being handed answers (use buddy:socratic-interview), or when they just want it built ("๊ทธ๋ฅ ๋ง๋ค์ด์ค",
A predict-then-verify + explain-back reading partner for UNFAMILIAR code, so the structure actually lodges in your head instead of evaporating with an AI summary. Instead of summarizing the code for you, it picks a spot, makes you PREDICT what it does (or who calls it, or what happens) via one AskUserQuestion, then reveals the real code so you learn from the gap โ and on the core spots it makes you EXPLAIN the module back in your own words until you can reconstruct it. The partner chooses the mode (predict vs explain-back) per spot by difficulty. You build the mental model; the partner never hands it over. TRIGGER: "grok this code", "help me understand this codebase", "read this code with me", "์ฝ๋ ์ดํด ๋์์ค", "๋ฏ์ ์ฝ๋ ํ์ ๊ฐ์ด ํด์ค", "์ด ์ฝ๋๋ฒ ์ด์ค ์ตํ๊ฒ ๋์์ค", "์ฝ๋ ๊ฐ์ด ์ฝ์", "์์ธกํ๋ฉฐ ์ฝ๋ ์ฝ๊ธฐ". DO NOT TRIGGER: the developer just wants a fast summary or answer ("๊ทธ๋ฅ ์์ฝํด์ค", "just tell me what this does"), needs to ship a fix right now, is doing a simple factual lookup (a signature, one line), or already understands the code and only needs
Socratic maieutic design interview, driven by structured multiple-choice questions. When a developer brings a vague idea, feature, or design problem, sharpen it into clear decisions by posing ONE focused AskUserQuestion at a time โ each option carrying its own reasoning ("X โ because โฆ") so the choice captures both the decision and its why. The options are a frame to think against, never a verdict; the choice, and the "Other" escape hatch, are always the developer's. The pen stays in their hand. TRIGGER: "socratic", "maieutic", "deep interview", "interview me to think", "help me decide by asking", "์ฐํ๋ฒ", "์ํฌ๋ผํ ์ค์์ผ๋ก ๋ฌผ์ด๋ด์ค", "์ค๊ณ ๊ฐ์ด ๊ณ ๋ฏผํด์ค", "์๊ฐ ์ ๋ฆฌ ์ธํฐ๋ทฐ". DO NOT TRIGGER: the developer just wants code or an answer fast ("just build it", "๊ทธ๋ฅ ์ง์ค"), wants a direct verdict without being questioned, is doing a simple factual lookup, or already has a detailed spec and only needs execution.
PRD-driven persistence loop. Reads .dt-handoff/<slug>/prd.json, drives each story RedโGreenโRefactor via test-engineer + executor, verifies acceptance criteria via the verifier agent on fresh evidence, runs reviewer approval, runs code-simplifier cleanup, re-verifies regression, and reports the stopping point. Auto commit/PR PROHIBITED. TRIGGER on "/ralph", "ralph", "๋๊น์ง ๊ฐ์ค", "keep going until done".
Consensus panel planning skill. Generates N risk-proportional candidate approaches via `architect` fan-out, then โ in one parallel wave โ independently evaluates them and has a SEPARATED `critic` adversarially attack each (authorโ attacker enforced by dispatch topology) and a distinct `critic` synthesize a ranked consensus ADR at `.dt-handoff/<slug>/plan.md`, marked `pending approval`. Panel width and attack depth scale to risk (LOW/MEDIUM/HIGH). TRIGGER on "/ralplan", "ralplan", "plan this", "ํฉ์ ๊ณํ". DO NOT TRIGGER for execution requests (those go to ralph / autopilot) or for free-form brainstorming.