autopilot-doc

name	autopilot-doc
description	Document strategy & draft pipeline — analyze-refs → strategy → review → draft → draft-review. 6 modes produce both strategy AND draft (markdown only). Inputs are discovered implicitly from .claude_reports/{analysis_project,research}/* (no --refs flag — pre-process external materials via /analyze-project --mode {paper|doc} first). All modes accept `--format-ref <path>` (universal flag — venue/journal/lab-specific guidelines, template, sample, or rebuttal-format file). No built-in presets (venues/years/journals differ). When omitted, agent auto-discovers a format spec inside analysis_project/doc/{matching}/formats/; review mode hard-fails if not found, other modes warn-and-fallback to a generic layout. presentation mode produces slide-by-slide markdown; PPTX export is NOT supported (use PowerPoint directly).
argument-hint	<task description> [--mode rebuttal|write|review|report|proposal|presentation] [--qa quick|light|standard|thorough] [--user-refine] [--no-clarify] [--from analyze|strategy|strategy-refine|draft|draft-refine|finalize]

산출물 폴더 컨벤션: SKILL_OUTPUT_CONVENTION.md (3-tier: T1 root / T2 named subdir / T3 _internal/). reviewer 로그는 _internal/strategy_reviews/·_internal/draft_reviews/. 버전 스냅샷은 _internal/versions/v{N}/strategy/, v{N}/draft/ (refine-doc의 _v{N}.md 형제 패턴은 폐기).

Language Rule

• Write user-facing output in Korean. (Material analysis results and pipeline_summary.md are written directly in the artifacts — no separate user output needed for those steps.)

Argument Parsing

Parse $ARGUMENTS for mode, flags, and task description:

--mode (optional, auto-inferred from query):

• rebuttal — Rebuttal strategy for reviewer comments
• write — Paper writing strategy (outline, positioning, contributions)
• review — Paper/document review (as reviewer: strengths/weaknesses/questions). Produces strategy + full review draft (OpenReview-ready).
• report — Technical report / white paper (findings, analysis, recommendations)
• proposal — Research or project proposal (problem, approach, plan, budget)
• presentation — Presentation strategy (story arc, slide structure, key messages) + slide-by-slide markdown draft. PPTX export is NOT performed by this pipeline (markdown-only); use PowerPoint manually with the lab template.

Auto-inference (mode 미지정 시):

• "리뷰 응답·rebuttal·reviewer comment" → rebuttal
• "발표·세미나·슬라이드·presentation" → presentation
• "리뷰·review·peer review·reviewer 입장" → review
• "보고서·report·기술 분석" → report
• "제안서·proposal·연구계획·grant" → proposal
• 그 외 / 명시적이지 않으면 → write (default for paper writing)
• 추론 결과를 한 줄로 사용자에게 통보 후 진행. 모호하면 Step 0 Scope Clarification에서 확인.

Note: survey mode is removed. For 학술/산업/시장 조사, use /autopilot-research --mode academic|technology|market first → autopilot-doc은 research/{topic}/ artifact를 implicit으로 자동 발견.

Input Discovery (implicit, no --refs flag) — --refs <folder>는 family에서 제거됨. 입력은 .claude_reports/ 하위 영속 산출물에서 자동 발견:

• analysis_project/paper/ — 보유 논문 분석 (autopilot-doc의 모든 모드에서 활용 가능)
• analysis_project/doc/{matching}/ — doc-creation 자료 (reviewer comments, format templates, samples). {matching}은 task description 키워드와 fuzzy match.
• research/{topic}/ — 외부 분야 조사 (autopilot-research가 만든 artifact). 마찬가지로 fuzzy match.
• analysis_project/code/ — 코드 컨텍스트 (proposal/report 모드에서 종종 인용)

mode별 필수 입력:

mode	필수 input	권장 input
rebuttal	analysis_project/doc/{matching}/reviewers/ (reviewer comments)	analysis_project/paper/ (원 paper 분석), analysis_project/doc/{matching}/formats/ (rebuttal guideline)
review	analysis_project/doc/{matching}/formats/ (review form)	analysis_project/paper/ (대상 paper 분석)
write	(없음)	analysis_project/paper/, analysis_project/doc/{matching}/formats/ (venue paper template), research/{topic}/ (분야 컨텍스트)
proposal, report, presentation	(없음)	mode 적합한 자료들 (위 패턴 응용)

매치 0: 사용자에게 안내 — "필요 자료를 analyze-project --mode {paper|doc} <folder> 로 먼저 사전 분석하세요" + 진행 여부 확인. 매치 다수: 후보 list 보여주고 선택 요청.

Implementation note (sub-skill / agent prompt 처리): 이하 본 SKILL.md 본문에서 {refs_folder}, --refs, Refs folder: 같은 표현은 _legacy 표기_이며 다음과 같이 해석·치환:

• 단일 {refs_folder} 변수 → discovered_inputs list를 newline-join한 문자열 (또는 콤마 join). Agent prompt에서 "Refs folder: {refs_folder}" 식 표현은 "Discovered inputs:\n - \n - \n ..." 로 expand.
• sub-skill 호출 (init-doc-strategy, refine-doc) → --inputs <comma-separated paths> 인수로 전달 (Step 2/3/5 참조).
• 이전 Decision Defaults / Safety Rules의 --refs 분기 → 모두 dead. Pre-flight Step 2 (Input Discovery)가 매치 0/다수 처리.

Mode별 자동 발견 입력 카테고리 (Pre-flight에서 결정):

• rebuttal: analysis_project/doc/{matching}/reviewers/ + analysis_project/paper/ (선택) + analysis_project/doc/{matching}/formats/ (선택)
• review: analysis_project/doc/{matching}/formats/ (REQUIRED) + analysis_project/paper/ (선택)
• write/proposal/report/presentation: 사용 가능한 analysis_project/paper/ + research/{topic}/ + analysis_project/doc/{matching}/formats/ 조합

본 SKILL.md를 점진적으로 정리하면서 legacy {refs_folder} 표현이 새 expand 형태로 대체될 예정.

--qa <level> — override QA intensity for the pipeline:

• --qa quick → fastest path: skip Step 3 (strategy refine) and Step 5 (draft refine) entirely + run a single sonnet quality reviewer pass at each review point with no re-invoke even if memos are added (memos are saved as audit trail, refine-doc is NOT invoked). --user-refine is silently ignored. fact-checker disabled.
• --qa light → 연구팀 review uses sonnet, single-pass review
• --qa standard → 연구팀 quality reviewer (opus) + 연구팀 fact-checker (sonnet, parallel) — fact-checker performs verbatim cards/PDFs 대조
• --qa thorough → 2× 연구팀 quality reviewers in parallel (opus, domain expert + methodology) + 연구팀 fact-checker (sonnet, parallel), cross-validation against all reference materials (default)
• If omitted, defaults to thorough.
• Why a separate fact-checker: quality reviewers focus on narrative/coverage/logic; fact-checker narrowly verifies citation/venue/year/metric/lineage against ground-truth sources (cards/PDFs). Sonnet is sufficient because fact-check is a matching task, not creative judgment.
• Propagation: Pass --qa <level> to init-doc-strategy and refine-doc as an argument flag.
• quick mode interactions: On --from strategy-refine or --from draft-refine, if frontmatter qa_level == quick, abort with: "qa_level=quick에서는 refine 단계가 스킵됩니다. --qa 을 다른 값으로 명시해 재개하세요."

--user-refine (boolean flag) — pause at refine points so the user can add their own <!-- memo: ... --> comments on top of 연구팀's memos before refine-doc runs.

Pause behavior: after 연구팀 writes memos at Step 3 (strategy review) or Step 5 (draft review), do NOT invoke refine-doc. Instead:

1. Update pipeline_state.yaml at {strategy_folder}/ with user_refine: true, paused_at_stage: <strategy-refine|draft-refine>.
2. Print to user (Korean) the memo file path and the resume command:

   연구팀 메모가 {ko_path}에 기록되었습니다.
   직접 메모를 추가한 뒤 다음 명령으로 재개하세요:
       /autopilot-doc --mode {mode} --from <strategy-refine|draft-refine> <strategy_folder>
   

3. Exit. Do NOT write pipeline_summary.md (pipeline is paused, not terminated).

If 연구팀 added no memos, the pause is skipped (nothing to refine).

--from <stage> — resume the pipeline at a specific stage. Stages:

• analyze — Step 1 (Material Analysis)
• strategy — Step 2 (init-doc-strategy)
• strategy-refine — Step 3 wrapper: 연구팀 review + (user memos if --user-refine) + refine-doc on the strategy
• draft — Step 4 (Draft Generation)
• draft-refine — Step 5 wrapper: 연구팀 review + (user memos if --user-refine) + refine-doc on the draft
• finalize — Step 6 (Pipeline Summary)

When resuming with --from, the positional argument should be either the artifact directory path or a fuzzy-matchable short name. The orchestrator resolves it via the same fuzzy lookup used by Plan Resolution in autopilot-code: ls -d .claude_reports/documents/*$ARG* 2>/dev/null. Read pipeline_state.yaml to recover mode, qa_level, discovered_inputs (list), user_refine. CLI flags override state file; missing flags inherit from state.

--format-ref <path> — universal flag. Path to a venue/journal/lab-specific format reference document. Available in every mode.

• No built-in presets. There is no single "openreview format" or "journal format" — even the same venue changes its review/rebuttal template year-to-year, and journals/labs each define their own. The user supplies the actual document for the target venue/round.
• Acceptable file types: .md, .txt, .pdf, .html, .docx (or any plain-text-ish format the agent can Read). Validated at Step 0 pre-flight.

What the file should contain (any subset — agent extracts what it can):

Mode	format-ref typical content
review	review template sections / rating axes (1-N with labels) / length limit / tone / submission portal layout
rebuttal	rebuttal length limit / allowed scope / sub-type indication (meta-reviewer-only one-shot vs reviewer-dialogue multi-round vs response-with-paper-revision) / submission window / examples of past-year rebuttals if available
write	venue paper template (e.g. NeurIPS 2026 LaTeX style) / page limits / section requirements / citation style / required disclosures
presentation	lab/venue slide template / time limits / required sections / branding rules / sample past presentations
proposal	grant body's required sections (NRF/NSF/internal) / page/word limits / required attachments / evaluation criteria
report	company/team report template / required sections / branding / audience expectations

Resolution order (every mode):

1. Explicit --format-ref <path> — agent reads it as authoritative format spec.
2. Auto-discovery in analysis_project/doc/{matching}/formats/ — if --format-ref is omitted, agent looks at analysis_project/doc/{matching}/formats/* (where {matching} was discovered by Input Discovery). The format extraction was already performed by analyze-project --mode doc. Acceptable extensions same as above.
   • 1 candidate found → use it, log to user: "format-ref auto-discovered: {path}".
   • 2+ candidates → ask user at Step 0 to pick one (or pass --format-ref <path> to specify).
   • 0 candidates → mode-specific fallback (next step).
3. Mode-specific fallback when neither explicit nor auto-discovered:

Mode	Behavior when no format-ref available
review	Hard fail at pre-flight — review mode cannot proceed without a venue review form. Abort with: "review mode requires either --format-ref <path> or a format file in analysis_project/doc/{matching}/formats/ (run /analyze-project --mode doc <folder> first). Venues differ year-to-year — no built-in presets."
rebuttal	Pre-flight prompt — ask user: (a) provide --format-ref now, or (b) declare format constraints inline in <task description> (length limit, sub-type, scope), or (c) opt into generic conference rebuttal layout (warn quality drop).
proposal	Warn-and-fallback to generic proposal layout. Recommend --format-ref <funding_body_template> for NRF/NSF/internal grants.
write	Warn-and-fallback to generic paper/article layout. Strong warning if target venue is academic — Suggest: "venue paper template (e.g. NeurIPS LaTeX style) significantly improves draft quality".
presentation	Warn-and-fallback to generic slide-by-slide markdown. Lab/venue slide templates improve fit but not blocking.
report	Warn-and-fallback to generic report layout. Suggest internal company template if applicable.

Sub-type information for rebuttal (meta-only / reviewer-dialogue / response-with-revision), section structure for review, page limits for write, etc. are all extracted from the format-ref file. No separate flags. If the file lacks the info, agent asks the user at Step 0 (within fallback prompt) or proceeds with documented assumptions.

The remaining text (after removing mode and flags) is the task description.

Note on presentation mode: This pipeline produces only the slide-by-slide markdown draft (draft/draft.md and draft/draft_ko.md). PPTX export is NOT supported because pandoc + Korean lab templates have unreliable compatibility (font/layout drift, OOXML strictness). The user converts markdown → PPT manually in PowerPoint using their lab template directly.

Decision Defaults (no autonomy gating)

The pipeline runs with sane defaults and only pauses on genuinely ambiguous or destructive situations.

Decision Point	Default Behavior
Confirm material analysis	Auto-proceed.
Missing refs folder	Always ask at pre-flight (mode-dependent).
No reviewer comments for rebuttal	Always ask at pre-flight.
Strategy review → memos added	Auto-refine (or pause for user-memo if --user-refine is set).
Draft review → memos added	Auto-refine (or pause for user-memo if --user-refine is set).
--format-ref missing	Auto-discover in analysis_project/doc/{matching}/formats/ (already classified by analyze-project --mode doc). If none, mode-specific fallback (review hard-fails; rebuttal prompts user; write/proposal/report/presentation warn-and-fallback to generic layout).
Reviewer guidelines absent in refs (review mode)	Use built-in spec only; inform user.
Scope Clarification triggered	Ask 2-4 questions; auto-proceed if --no-clarify.

Logging: When the pipeline pauses (missing required input, 0 search results, or --user-refine), record the event for the Decision Points table in pipeline_summary.md. Auto-decisions are not individually logged.

pipeline_state.yaml

Written/updated at {strategy_folder}/pipeline_state.yaml after each completed stage. Used by --from resume:

pipeline: autopilot-doc
mode: presentation
qa_level: thorough
user_refine: true
discovered_inputs:                    # list of paths discovered by Pre-flight Step 2 (Input Discovery)
  - <path-to-analysis_project/paper-or-doc-or-research-artifact>
  - ...
format_ref: <path or null>          # universal — explicit flag, auto-discovered, or null after fallback
format_ref_source: <explicit|auto-discovered|user-supplied-at-prompt|fallback-generic>
clarified_intent: <string or null>    # captured by Step 0 Scope Clarification, used on resume
last_completed_stage: strategy        # one of: clarify, analyze, strategy, strategy-refine, draft, draft-refine, finalize
paused_at_stage: strategy-refine      # set only when --user-refine triggered a pause
artifact_dir: <abs path>

CLI flags on resume override stored values. After the pause is consumed (refine completes), clear paused_at_stage and update last_completed_stage.

Input Sources Convention

External materials must be pre-processed into .claude_reports/ before invoking autopilot-doc. The pipeline reads from these persistent sources only — no --refs flag, no ad-hoc folder paths.

Input type	Pre-processing skill	Output location
Academic papers (PDFs)	/analyze-project --mode paper	analysis_project/paper/
Reviewer comments / format templates / past samples / mixed doc materials	/analyze-project --mode doc <folder>	analysis_project/doc/{name}/
External field research	/autopilot-research <topic>	research/{topic}/
Codebase context (proposal/report 모드에서 언급용)	/analyze-project --mode code	analysis_project/code/

On invocation, autopilot-doc runs Input Discovery (Pre-flight Step 2) — fuzzy match task description vs above persistent sources — and gathers discovered_inputs paths to pass to sub-skills. For rebuttal mode, fails with clear message if no reviewer materials match.

Artifact Structure

All outputs go to:

.claude_reports/documents/{YYYY-MM-DD}_{short-name}/
├─ pipeline_summary.md       (T1 — entry/index + integrated history)
├─ draft/                    (T1 — generated for all 6 modes; latest only)
│  ├─ draft.md              (English draft; for presentation: slide-by-slide markdown)
│  └─ draft_ko.md           (Korean draft)
├─ strategy/                 (T2 — latest only)
│  ├─ strategy.md           (English strategy document)
│  └─ strategy_ko.md        (Korean strategy document)
├─ analysis/                 (T2)
│  ├─ reviewer_analysis.md   (rebuttal: per-reviewer breakdown)
│  ├─ ref_analysis.md        (reference material analysis)
│  └─ material_index.md      (inventory of all input materials)
└─ _internal/                (T3 — audit / reviews / version snapshots)
   ├─ strategy_reviews/      (QA and 연구팀 strategy reviews)
   ├─ draft_reviews/         (QA and 연구팀 draft reviews)
   └─ versions/              (autopilot-refine snapshots)
      ├─ v1/strategy/, draft/
      └─ v{N}/...

Pipeline

Pre-flight Validation [ALL modes — runs first, before any work]

Validate mode-specific required inputs. If any check fails, abort immediately with a clear error message — do NOT create the artifact directory or invoke any sub-skills/agents.

Universal checks (all modes):

1. Mode is one of the 6 supported modes (rebuttal / write / review / report / proposal / presentation) — explicit --mode 또는 auto-inference. Otherwise abort: "Unknown mode: {mode}. Supported: ...".
2. Input Discovery (replacing old --refs check): run fuzzy match on task description vs .claude_reports/analysis_project/{paper,doc}/* and .claude_reports/research/*. Per mode:
   • rebuttal: at least one match in analysis_project/doc/*/reviewers/ REQUIRED. If none → abort with: "rebuttal mode needs reviewer comments. Run /analyze-project --mode doc <folder> first to materialize them."
   • review: at least one match in analysis_project/doc/*/formats/ REQUIRED. If none → abort with similar message for review form.
   • write/proposal/report/presentation: no hard requirement, but warn if no matches at all and ask user to confirm.
   • Stash discovered paths into orchestrator context as {discovered_inputs} for downstream Steps.

Mode-specific checks:

Universal --format-ref resolution (runs before mode-specific checks):

1. If --format-ref <path> explicit → validate path exists + extension in {.md,.txt,.pdf,.html,.docx}. Otherwise abort.
2. If omitted → auto-discover in analysis_project/doc/{matching}/formats/ (already classified by analyze-project --mode doc):
   • 1 candidate → use it; log "format-ref auto-discovered: {path}".
   • 2+ candidates → ask user at Step 0 which to use, or to pass --format-ref explicitly.
   • 0 candidates → mode-specific fallback below.

Mode-specific pre-flight (after universal resolution):

• review mode — format-ref is REQUIRED.

  • If still no format-ref after auto-discovery → abort with: "review mode requires either --format-ref <path> or a format file in analysis_project/doc/{matching}/formats/ (run /analyze-project --mode doc <folder> first to extract). Venues differ year-to-year — no built-in presets. Acceptable file types: .md/.txt/.pdf/.html/.docx."

• rebuttal mode — two checks:

  • refs folder must contain at least one reviewer-comment file (txt/md/pdf with reviewer-style content detected by filename or content scan). If none found, ask the user before proceeding.
  • format-ref absent (no flag, no auto-discovery hit) → prompt user at Step 0: "(a) provide --format-ref now / (b) declare format constraints (length, sub-type, scope) inline in / (c) opt into generic conference rebuttal layout (warns quality drop)". Sub-type info (meta-only / reviewer-dialogue / response-with-revision) is extracted from the format-ref file or stated in task description — no separate flag.

• presentation mode — format-ref optional. If absent, fallback to generic slide-by-slide markdown layout (warning logged).

• proposal / report / write modes — format-ref optional. If absent, fallback to generic mode-specific layout. For write targeting an academic venue (detected from task description or refs/), strongly recommend supplying the venue's paper template.

Abort behavior:

• Print the error message in Korean to the user.
• Do NOT call mkdir, do NOT invoke any sub-skill, do NOT write pipeline_summary.md.
• Exit with status: aborted (pre-flight).

After all pre-flight checks pass: create artifact_dir and proceed to Step 0.

Step 0: Scope Clarification (사전 조율) — skipped if --no-clarify

Purpose: Catch ambiguous queries before launching the pipeline. autopilot-doc 산출물 품질은 task 명확도에 비례하므로, 모호한 입력은 30% signal·70% noise를 만든다.

Trigger conditions (any one matches → run clarification):

• Mode auto-inference 신뢰도 낮음 (키워드 매치 약함, 또는 multi-match)
• Task description < 15 words AND no specific deliverable hint
• Mode가 review인데 venue/length/style 미명시
• Mode가 presentation인데 청중·시간 미명시
• Mode가 proposal인데 grant body·deadline·예산 범위 미명시

Action: 메인 Claude가 mode-aware 2-4개 sharp question을 던진다. 사용자 답변을 task description에 통합 후 Step 1 진행.

Mode-specific question seed:

• write / report / proposal: 청중, 길이/페이지 제한, 강조 포인트, deadline
• presentation: 청중 (전공자/비전공자/임원), 시간 (30/45/60min), 핵심 메시지 1개
• review: venue / 리뷰 가이드라인 / 점수 체계
• rebuttal: rebuttal 길이 제한, 추가 실험 가능 여부, 톤 (defensive vs concessive)

Skip 조건:

• --no-clarify 명시
• task description이 충분히 구체적 (12+ words + concrete deliverable + constraints)
• --from <stage> 재개 (기존 pipeline_state.yaml에 이미 정보 있음)

Output: 사용자 답변을 통합한 refined task description을 메모리에 저장 + pipeline_state.yaml의 clarified_intent 필드에 기록.

Step 1: Material Analysis

Read and catalog all materials from refs folder.

1. Inventory: List all files with brief descriptions. Write to analysis/material_index.md.
2. Analyze by mode:
   • rebuttal: Parse reviewer comments → analysis/reviewer_analysis.md (per-reviewer, per-point breakdown with severity classification)
   • write: Analyze reference papers → analysis/ref_analysis.md (methods, gaps, positioning opportunities)
   • review: Analyze target paper/document → analysis/ref_analysis.md (methodology assessment, quality analysis)
   • report: Analyze source data/papers → analysis/ref_analysis.md (findings, evidence assessment, data quality)
   • proposal: Analyze related work and context → analysis/ref_analysis.md (prior art, feasibility evidence, competitive landscape)
   • presentation: Analyze source document/paper → analysis/ref_analysis.md (key messages, audience analysis, narrative structure)
3. Read PDF files using the Read tool. For large PDFs (>10 pages), read in page ranges.
4. Present the analysis summary briefly and auto-proceed to Step 2 — no confirmation required.

Step 2: init-doc-strategy

Invoke Skill: init-doc-strategy with args: <mode> --inputs <comma-separated-discovered-paths> --output <artifact-dir> <task description>. <discovered-paths>는 Pre-flight Step 2 (Input Discovery)가 발견한 analysis_project/{paper,doc}/..., research/{topic}/ 경로 list (콤마 join). 매치 0이면 Pre-flight에서 이미 abort/warn 처리됨. Wait for completion.

Step 3: Strategy Review (연구팀 as domain expert)

1. Resolve strategy paths:

   • strategy_folder = .claude_reports/documents/{YYYY-MM-DD}_{short-name}/
   • en_strategy_path = {strategy_folder}/strategy/strategy.md
   • ko_strategy_path = {strategy_folder}/strategy/strategy_ko.md

2. Invoke reviewers based on --qa level. Quality reviewer(s) and fact-checker run in parallel at standard+:

   quick — Single 연구팀 quality reviewer (sonnet, spot-check only):

   • One-pass review. Memos may be added but refine-doc is NOT invoked at Step 3 (see step 3 below).
   • Review log: {strategy_folder}/_internal/strategy_reviews/research_review.md

   light — Single 연구팀 quality reviewer (sonnet):

   • One-pass review focusing on critical issues only.
   • Review log: {strategy_folder}/_internal/strategy_reviews/research_review.md

   standard — 1× 연구팀 quality reviewer (opus) + 1× 연구팀 fact-checker (sonnet, parallel):

   • Quality review log: {strategy_folder}/_internal/strategy_reviews/research_review_quality.md
   • Fact-check log: {strategy_folder}/_internal/strategy_reviews/research_review_factcheck.md

   thorough (default) — 2× 연구팀 quality reviewers (opus, parallel) + 1× 연구팀 fact-checker (sonnet, parallel):

   • Quality Reviewer A (Domain Expert): Cross-checks strategy against reference materials, domain conventions (academic venues for paper modes: NeurIPS, ICML, ICLR, ICASSP, Interspeech, T-ASLP; industry standards for report/proposal/presentation modes), and completeness of coverage.
     • Review log: {strategy_folder}/_internal/strategy_reviews/research_review_domain.md
   • Quality Reviewer B (Methodology Reviewer): Evaluates logical consistency, persuasiveness of arguments, experimental design soundness, and identifies potential weaknesses an adversarial reviewer would exploit.
     • Review log: {strategy_folder}/_internal/strategy_reviews/research_review_methodology.md
   • Fact-checker (sonnet, parallel): Verbatim cross-check of citation/venue/year/metric/lineage against cards/PDFs.
     • Review log: {strategy_folder}/_internal/strategy_reviews/research_review_factcheck.md
   • All reviewers write <!-- memo: ... --> comments in the Korean strategy.
   • After all complete, merge memos and deduplicate.

   Quality reviewer prompt (light/standard/thorough A & B):

   Review this document strategy as the user's domain expert proxy — _quality / cohesion / coverage_ focus.
   Mode: {mode} | KO strategy: {ko_strategy_path} | EN strategy: {en_strategy_path}
   Analysis: {strategy_folder}/analysis/ | Refs: {refs_folder} | Log: {review_log_path}
   
   Cross-check: actual refs/reviewer comments, domain conventions,
   logical consistency, completeness (any missed reviewer points or gaps?).
   Do NOT verify individual fact citations (model venue/year/metric) — that's the fact-checker's role at standard+.
   Write memos as `<!-- memo: ... -->` in the Korean strategy.
   Write a structured review log to the log file.
   Return a summary of memos added (or "no issues found").
   

   Fact-checker prompt (sonnet, parallel — standard/thorough only):

   You are a fact-check focused reviewer — NOT narrative quality.
   Mode: {mode} | KO strategy: {ko_strategy_path} | Refs: {refs_folder} | Log: {fact_log_path}
   
   For every domain claim in the strategy (citation / model name / venue / year /
   metric / dataset / lineage / classification), open the corresponding ground-truth
   source and verbatim compare:
   - Paper cards: {refs_folder}/cards/*.md (if exists; this is single source of truth)
   - Reference PDFs: {refs_folder}/*.pdf (only if cards lack the specific fact)
   - Reviewer comments (rebuttal mode): {strategy_folder}/analysis/reviewer_analysis.md
   
   Do NOT comment on completeness, narrative arc, or strategic soundness — that's the quality reviewer's job.
   Stay narrowly on fact verification. Cost-aware mode (sonnet): table-only output. Limit to ~30 most material claims.
   
   Output the review log as a single table:
   | Section | Claim in strategy | Source (file:line or section) | Match (✅/❌) | Severity (🔴/🟡) |
   
   For 🔴/🟡 mismatches, also write `<!-- memo: [FACT] section X — claim Y conflicts with source Z -->` in the Korean strategy.
   Return ONLY path + one-line verdict.
   

3. If memos were added:

   • qa_level == quick short-circuit: do NOT invoke refine-doc. Memos remain in the strategy as audit trail (no edits applied). Log to pipeline_summary Decision Points: Step 3 | strategy refine skipped (qa=quick) | auto | proceed to Step 4. Skip to Step 4.
   • --user-refine pause: if the flag is set, update pipeline_state.yaml (user_refine: true, paused_at_stage: strategy-refine), print the resume command (/autopilot-doc --mode {mode} --from strategy-refine {strategy_folder}), and exit. Do NOT invoke refine-doc.
   • Otherwise: invoke Skill refine-doc with the Korean strategy path as args.

4. If no memos: Skip to Step 4. (When resumed via --from strategy-refine, the orchestrator skips the 연구팀 review and runs refine-doc directly using the pre-existing memos.)

Step 4: Draft Generation

Applicable modes: rebuttal, write, report, proposal, review, presentation. (All 6 modes generate drafts.)

1. Verify strategy is finalized: {strategy_folder}/strategy/strategy.md exists and has no ## 미해결 이슈 section (or issues are acceptable).
2. Invoke the research-team (연구팀) agent as a subagent:

Draft generation mode. Generate a document draft based on the finalized strategy.

Mode: {mode}
Task: {task description}
Strategy (EN): {en_strategy_path}
Strategy (KO): {ko_strategy_path}
Analysis directory: {strategy_folder}/analysis/
Reference materials: {refs_folder}
Save English draft to: {strategy_folder}/draft/draft.md
Save Korean draft to: {strategy_folder}/draft/draft_ko.md

Read the strategy document and all analysis files. Generate a complete first draft following the mode-specific structure below. The draft should be a working document ready for user editing — not a summary of the strategy.

## Tone Propagation (modes: report / proposal / presentation)

**FIRST**, read the strategy frontmatter `tone` field:
- If `tone: administrative` — apply administrative-tone constraints to the **entire draft** (slide titles, bullets, conclusion, visual placeholders). Specifically:
  - **AVOID**: marketing superlatives ("genuinely novel", "sole occupied axis", "global rights asset", "world-first", "compelling contribution"), "X strengths summary" framing, "core message" + "Hook → Call-to-Action" arc, heroic asks ("Approve to secure as global asset"), decision-options box (approve/conditional/hold), animated narrative voice
  - **PREFER**: simple fact lists, status updates, neutral reporter stance, calm review request ("검토 부탁드립니다" / "kindly request the committee's review")
  - Conclusion slide: replace "Key messages + Call-to-Action" with **"Presentation summary + review request"**; remove "X strengths" enumeration in favor of plain fact recap
  - Speaker stance: **neutral reporter, not advocate**. The speaker (often a student or researcher) is reporting upward to decision-makers, not pitching to peers
- If `tone: default` or absent — existing pitch-deck patterns apply (Hook, Core Message, Story Arc, Call-to-Action, persuasive framing)

This propagation is mandatory: a `tone: administrative` strategy with a heroic-pitch draft is a critical mismatch and must be reworked.

## Mode-Specific Draft Structure

### rebuttal
- Frontmatter: type, venue, status: draft, date
- Per-reviewer response sections following the strategy's priority matrix
- Each response: acknowledgment → core argument → evidence → conclusion
- Tone calibrated per the strategy's tone guidelines
- Additional experiments section with preliminary descriptions
- Revision summary table

### write
- Frontmatter: type, venue, status: draft, date
- Full paper outline with section drafts:
  - Abstract (structured: background → gap → method → results → impact)
  - Introduction (hook → context → gap → contribution → outline)
  - Related Work (organized by strategy's framing)
  - Method (following strategy's outline, with placeholder equations)
  - Experiments (setup → results → ablation, with table skeletons)
  - Conclusion
- Figure/table placeholders with captions

### report
- Frontmatter: type, status: draft, date
- Executive Summary
- Introduction / Background
- Methodology / Approach
- Findings / Analysis (with data tables, charts description)
- Discussion
- Recommendations (prioritized, actionable)
- Appendices (if needed)

### proposal
- Frontmatter: type, status: draft, date
- Executive Summary
- Problem Statement / Motivation
- Proposed Approach / Technical Plan
- Preliminary Results / Feasibility Evidence
- Timeline & Milestones
- Resource Requirements / Budget (if applicable)
- Expected Outcomes / Impact
- Risk Assessment

### review
Adapt the section structure to the file at `--format-ref` (read it first). No built-in presets — extract the venue's required sections / rating axes / length limits from the user-supplied format-ref. If extra reviewer guidelines exist in refs/, layer them on top.

**Frontmatter** (always): type, venue, paper_title, status: draft, date, format_ref (path to user-supplied format spec)

**Procedure**:

1. Read the file at `--format-ref` first. Extract: required sections, rating axes (with score scales 1-N and meanings), length limits, tone/style guidelines, submission portal layout.
2. If the format-ref is a venue's reviewer guidelines PDF/doc, prefer its exact section names verbatim. If it's a sample review, infer the structure.
3. Layer any additional reviewer guidelines from `analysis_project/doc/{matching}/formats/` on top.
4. Produce a draft that satisfies every required section from the format-ref.

**Common patterns** (reference only — the actual structure must come from format-ref, not from these):

- _OpenReview-family_ (NeurIPS, ICML, ICLR, AAAI variants): Summary / Strengths / Weaknesses / numeric ratings (Soundness, Presentation, Significance, Originality on 1-4 or 1-5) / Questions / Limitations / Overall Recommendation + Confidence
- _ACL ARR_: Paper Summary / Strengths / Weaknesses / Comments+Typos / Soundness, Excitement, Reproducibility (1-5) / Ethical Concerns
- _IEEE conference_ (ICASSP, INTERSPEECH): Brief Summary / Strengths / Weaknesses / Detailed Comments / Recommendation (Accept/Reject scale) / Confidence
- _Journal_ (T-ASLP, JASA, TPAMI, etc.): Significance / Technical Quality / Clarity / Recommendation (Accept/Minor Revision/Major Revision/Reject) / Per-section comments

These are starting hints only. Always follow the format-ref file's actual specification — venue templates change year-to-year.

### presentation
Generate a **PPT cheatsheet markdown** — single file, optimized for human reading and slide-by-slide copy/paste into PowerPoint. **NOT a pandoc conversion target**. Avoid pandoc-specific syntax (`::: notes`, `:::: {.columns}`, YAML frontmatter for auto-title generation).

**Slide Format Conventions** (mandatory — derived from user feedback to prevent revision loops):

1. **Chapter visualization in slide headers** — every body slide's heading: `## Slide N — [Ch.N 챕터명] (sub.번호) 슬라이드 제목`. Chapter-transition slides marked with `— 시작` / `— start`. Each slide has a `**챕터**: N. 챕터명 (M장 중 K번째)` meta line below the title.

2. **Visual placeholder must include chapter band** — every body slide's `**시각자료**:` block first line: `- **상단 헤더 띠**: "N. 챕터명"` (per Korean industry-academia format-ref). Chapter-transition slides additionally specify "Ch.X와 색상/strength를 다르게 — 챕터 전환 시각 신호".

3. **Concrete visual placeholders** — NO vague terms like "X 카드", "적절한 도식", "comparison chart". Every visual specifies (a) diagram type + (b) component list + (c) layout/color hints. Example: ❌ "학회 위상 카드" → ✅ "NeurIPS/ICLR/ICML 3-row table (h5-index 컬럼 + acceptance-rate 컬럼)".

4. **Table column header clarity** — NO ambiguous headers like "비교 1위" or "vs ours". Use full noun phrases with clear semantic units. If needed, add a 1-line column-meaning footnote above the table.

5. **Foreign-language quote → Korean keyword gloss** (mandatory for non-AI audiences) — every English quote (paper review citation, technical term, model description) gets a Korean appeal-commentary box directly below:

"English quote..." — Source

📌 핵심 키워드 — "X": 한국어 풀이 1문장 (청중 친화 어필 메시지)

6. **Speaker notes default = empty** — do NOT auto-fill speaker notes in the initial draft. Wait for explicit user request as a separate post-polish step. Reason: speaker notes drift with slide-content edits; auto-fill wastes regeneration cost during iterative refinement.

7. **No body-bullet ↔ visual redundancy** — the same fact should NOT appear in both body bullets AND visual placeholder. Body bullets = "what the speaker says"; visual = "what the audience sees at-a-glance". If redundant, simplify one of the two.

8. **Slide-number consistency on insertion/deletion** — when inserting/removing/renumbering a slide, update ALL of the following in the same edit pass:
- (a) All subsequent slide numbers (`Slide N+1`, `Slide N+2`, ...)
- (b) Contents slide's chapter slide-counts ("Ch.N (M장)")
- (c) CHANGELOG entry at file top
- (d) Time-budget line in the top-of-file guide
- (e) Cross-references in other slides ("Slide M의 ...")
- (f) Chapter meta lines ("M장 중 K번째")

**Top-of-file guide** (mandatory header before any slides):

```markdown
# {발표 제목} — Seminar Slide Deck

> **사용 가이드**: 본 markdown은 PPT 복사·붙여넣기용 단일 파일이다. 각 슬라이드는 `---`로 분리되어 있으며, 슬라이드 번호·제목·bullet·시각자료·Speaker note 순서로 구성된다.
>
> - **총 슬라이드 수**: **N main + M backup = total**
> - **시간 분배 ({X}분 기준)**: Opening / Ch.0 / Ch.1 / ... 분 단위 명시
> - **청중 baseline**: 한 줄로 청중 특성과 작성 톤 (약어 풀어쓰기 / 직관 비유 / 수식 최소 등)
> - **설계 의도**: 챕터 구성·narrative arc 한 단락

슬라이드 단위 형식 (모든 main + backup 슬라이드):

---

## Slide N — {짧은 슬라이드 제목}

**제목**: {실제 슬라이드에 들어갈 제목 문구 (한국어 또는 본인이 쓰는 발표 언어)}

**부제** (선택): {부제 문구 — 첫 슬라이드 또는 챕터 디바이더에 한정}

- 본문 bullet 1 (개념/이름/수치 위주, 간결하게)
- 본문 bullet 2
- 본문 bullet 3 (보통 3-5개)

| 표가 더 적합한 경우 | 이렇게 markdown 표 |
|---|---|
| 모델 A | 수치 |
| 모델 B | 수치 |

**시각자료**:
- 좌측 1/2 (또는 메인): {도식·차트 설명}
- 우측 1/2 (또는 보조): {보조 시각}
- 또는 전체 화면: {풀 페이지 도식 설명}

**Speaker note**:
1. {발화 1 — 슬라이드 본문 보충, 직관 풀이, 비유, 일화}
2. {발화 2 — 다음 슬라이드/챕터로 가는 transition}
3. {발화 3 — 청중 질문 예상 시 짧은 답변 메모, 선택}

**Citation** (선택): [Author Year, Venue](cards/{file}.md) — 정확한 paper card를 가리키는 인라인 링크

구조 요건:

• 표지 (Slide 1) — 제목 + 부제 + 발표자/소속 + 날짜 + 발표 자료 출처 한 줄
• 목차 (Slide 2) — 챕터별 슬라이드 수와 한 줄 설명
• 챕터 디바이더 — ## Slide N — Ch.X 제목 형식. 슬라이드 본문은 챕터 의도/시기 한두 줄. 별도 슬라이드 카운트에 포함.
• 본문 슬라이드 — 위 슬라이드 단위 형식
• 챕터 마무리 (선택) — Ch.X 정리 + Ch.X+1 transition. 인지 부담 분산용
• Conclusion — Take-home 5 / Open Problems / 한 페이지 요약 / Q&A / Thank you
• Backup — ## Slide BN — Backup: 제목 형식. 메인 흐름 끝난 뒤 배치
• References (선택) — 마지막에 핵심 인용 정리

작성 톤:

• 본문 bullet은 키워드 + 수치 + 모델명 위주. 풀 문장 지양 (그건 speaker note에).
• 약어는 첫 등장 시 풀어쓰기: Speech Enhancement (SE), NFE (Number of Function Evaluations) 등.
• Citation은 paper card markdown 링크로 ([Author Year](../../research/{topic}/cards/{file}.md) 또는 같은 artifact_dir 내 cards/).

Quality:

• 모든 본문 슬라이드에 Speaker note 필수 (≥80% — 기술 비중 낮은 표지·인사 슬라이드 제외).
• 모든 슬라이드에 시각자료 placeholder (텍스트만으로 끝나는 슬라이드는 cheatsheet로서 약함).
• 시각자료 설명은 PPT에서 그릴 수 있을 만큼 구체적으로 (예: "5-stage timeline 가로 막대, 색상 5개" 같은 수준).
• Strategy doc의 슬라이드 outline을 그대로 매핑 (총 슬라이드 수와 챕터 시간 분배 일치).

Quality Requirements

• Every claim must trace back to a specific reference in the refs folder or analysis.
• Do NOT fabricate citations, data, or results.
• Mark uncertain or placeholder content with [TODO: ...].
• Mode-specific completeness criteria:
  • rebuttal: 90%+ — every reviewer point MUST have a drafted response (hard constraint). Missing a point is a critical error.
  • write/report/proposal: 70-80% — all sections with substantive content, no heading-only sections.
  • review: 80%+ — every required section per the --format-ref file must be filled with concrete claims. Strengths/weaknesses must reference specific paper sections/figures/tables. Score justifications are mandatory.
  • presentation: 70-80% — every slide has 제목/부제(선택)/bullets/시각자료/Speaker note 5개 슬롯이 채워짐 (시각자료가 텍스트만 있는 슬라이드에 빠지면 cheatsheet 가치 손상). Speaker notes ≥80% of content slides. 슬라이드 카운트는 strategy outline과 ±10% 이내. --- 구분자가 모든 슬라이드 사이에 있는지 확인.

Write both files directly. Return ONLY the file paths and a 3-5 line Korean summary.

3. **IMPORTANT**: Do NOT read, re-write, or duplicate the draft files yourself. The agent writes them directly.

### Step 5: Draft Review (연구팀 as QA)
**Applicable modes**: rebuttal, write, report, proposal, review, presentation. (All 6 modes that generated drafts.)

1. Resolve draft paths:
   - `en_draft_path` = `{strategy_folder}/draft/draft.md`
   - `ko_draft_path` = `{strategy_folder}/draft/draft_ko.md`

2. Invoke reviewers based on `--qa` level (same scaling as Step 3). **Quality reviewer(s) and fact-checker run in parallel** at standard+:

   **`quick`** — Single 연구팀 quality reviewer (sonnet, spot-check only):
   - One-pass review. Memos may be added but refine-doc is NOT invoked at Step 5 (see step 3 below).
   - Review log: `{strategy_folder}/_internal/draft_reviews/draft_review.md`

   **`light`** — Single 연구팀 quality reviewer (sonnet):
   - One-pass review focusing on critical issues only.
   - Review log: `{strategy_folder}/_internal/draft_reviews/draft_review.md`

   **`standard`** — 1× 연구팀 quality reviewer (opus) + 1× 연구팀 fact-checker (sonnet, parallel):
   - Quality review log: `{strategy_folder}/_internal/draft_reviews/draft_review_quality.md`
   - Fact-check log: `{strategy_folder}/_internal/draft_reviews/draft_review_factcheck.md`

   **`thorough`** — 2× 연구팀 quality reviewers (opus, parallel) + 1× 연구팀 fact-checker (sonnet, parallel):
   - **Quality Reviewer A (Content Expert)**: Cross-checks draft against strategy, verifies all strategy points are addressed, checks high-level factual coherence.
     - Review log: `{strategy_folder}/_internal/draft_reviews/draft_review_content.md`
   - **Quality Reviewer B (Writing Quality)**: Evaluates writing quality, logical flow, completeness, identifies gaps and weak arguments.
     - Review log: `{strategy_folder}/_internal/draft_reviews/draft_review_quality.md`
   - **Fact-checker (sonnet, parallel)**: Verbatim cross-check of citation/venue/year/metric/lineage against cards/PDFs — _independent_ of strategy/quality concerns.
     - Review log: `{strategy_folder}/_internal/draft_reviews/draft_review_factcheck.md`
   - All reviewers write `<!-- memo: ... -->` comments in the Korean draft.
   - After all complete, merge memos and deduplicate.

   **Quality reviewer prompt** (light/standard/thorough A & B):

Review this document draft as the user's domain expert proxy — strategy coverage / writing quality / logic focus. Mode: {mode} | KO draft: {ko_draft_path} | EN draft: {en_draft_path} Strategy: {en_strategy_path} | Analysis: {strategy_folder}/analysis/ | Refs: {refs_folder} Log: {review_log_path}

Cross-check: strategy coverage (all points addressed?), logical flow, writing quality, completeness, [TODO] items. For rebuttal: verify every reviewer point has a response. Do NOT individually verify each fact citation (model venue/year/metric) — that's the fact-checker's role at standard+. Write memos as <!-- memo: ... --> in the Korean draft. Write a structured review log to the log file. Return a summary of memos added (or "no issues found").

**Fact-checker prompt** (sonnet, parallel — standard/thorough only):

You are a fact-check focused reviewer — NOT narrative quality. Mode: {mode} | KO draft: {ko_draft_path} | Refs: {refs_folder} | Log: {fact_log_path}

For every domain claim in the draft (citation / model name / venue / year / metric / dataset / lineage / classification), open the corresponding ground-truth source and verbatim compare:

• Paper cards: {refs_folder}/cards/*.md (if exists; this is single source of truth)
• Reference PDFs: {refs_folder}/*.pdf (only if cards lack the specific fact)
• Strategy: {en_strategy_path} (only to confirm consistent claim, not as primary source)

Do NOT comment on writing quality, narrative arc, or strategy coverage — that's the quality reviewer's job. Stay narrowly on fact verification. Cost-aware mode (sonnet): table-only output. Limit to ~30 most material claims.

Output the review log as a single table: | Slide/Section | Claim in draft | Source (file:line or section) | Match (✅/❌) | Severity (🔴/🟡) |

For 🔴/🟡 mismatches, also write <!-- memo: [FACT] slide X — claim Y conflicts with source Z --> in the Korean draft. Return ONLY path + one-line verdict.

3. If memos were added:
- **`qa_level == quick` short-circuit**: do NOT invoke refine-doc. Memos remain in the draft as audit trail (no edits applied). Log to pipeline_summary Decision Points: `Step 5 | draft refine skipped (qa=quick) | auto | proceed to Step 6`. Skip to Step 6.
- **`--user-refine` pause**: if the flag is set, update `pipeline_state.yaml` (`user_refine: true`, `paused_at_stage: draft-refine`), print the resume command (`/autopilot-doc --mode {mode} --from draft-refine {strategy_folder}`), and exit. Do NOT invoke refine-doc.
- Otherwise: invoke Skill `refine-doc` with the Korean draft path as args.
- Note: refine-doc handles draft paths (draft/draft.md ↔ draft/draft_ko.md) via auto-detection.
4. If no memos: Skip to Step 6. (When resumed via `--from draft-refine`, run refine-doc directly on the pre-existing memos.)

### Step 6: Pipeline Summary
**Always write** `{strategy_folder}/pipeline_summary.md` before reporting to the user.

```markdown
# Document Strategy Pipeline Summary: {task name}

- **Date**: {YYYY-MM-DD} | **Mode**: {mode} | **Format-ref**: {format_ref or "fallback-generic"} ({format_ref_source}) | **Status**: done / reviewed / draft
- **User-Refine**: {true | false}
- **Refs folder**: {refs_folder}

## Process Log
| Step | Action | Result | Notes |
|---|---|---|---|
| 0 | Scope Clarification | clarified / skipped | {questions asked or "--no-clarify"} |
| 1 | Material Analysis | completed | {N} files |
| 2 | init-doc-strategy | created | {strategy path} |
| 3 | Strategy Review (연구팀) | memos added / no issues | {memo count} |
| 3b | refine-doc | refined / skipped | |
| 4 | Draft Generation | created | {draft path} |
| 5 | Draft Review (연구팀) | memos added / no issues | {memo count} |
| 5b | refine-doc (draft) | refined / skipped | |

## Artifacts
- Strategy (EN/KO): {en_path} / {ko_path}
- Draft (EN/KO): {draft_en_path} / {draft_ko_path}
- Analysis: {reviewer_analysis or ref_analysis path}
- Material Index: {path} | Strategy Review: {path} | Draft Review: {path}

## Decision Points
| Step | Decision | User Response | Action Taken |
|---|---|---|---|
| (filled from orchestrator's in-memory decision log) |

When writing pipeline_summary.md, populate the Decision Points table from the in-memory decision records. If no decisions were recorded (clean run with no --user-refine, no missing inputs), write: | - | No pause points triggered | - | - |.

Then report to the user:

• Strategy file paths + 2-3 line summary of the strategy.
• Draft file paths + 2-3 line summary of the draft.
• For presentation mode: remind the user that PPTX export is manual — they should open the markdown draft and copy slide content into PowerPoint with their lab template.
• For review mode: confirm the --format-ref file used and any venue-specific adaptations from refs/. No built-in presets.

Safety Rules

• Do NOT fabricate citations or invent results — only reference materials actually present in the refs folder.
• The draft is a working first draft for user editing, NOT a final document. Mark uncertain content with [TODO: ...].
• For rebuttal mode: ensure EVERY reviewer point is addressed — missing a point is a critical error.
• For review mode: scores must be justified with concrete evidence; never fabricate scores without backing in the paper text. --format-ref <path> (explicit or auto-discovered) is mandatory — pre-flight aborts otherwise.
• For rebuttal mode: rebuttal sub-type (meta-only / reviewer-dialogue / response-with-revision) must be derivable from format-ref content OR task description by Step 1. Strategy and tone differ across sub-types — if neither source provides it, Step 0 prompt asks the user to declare.
• For all other modes: format-ref is optional but improves quality significantly when supplied. The agent should note the format-ref source in the strategy frontmatter so future refine-doc rounds know what to honor.
• For presentation mode: never insert real figures/images automatically — describe visuals in the **시각자료**: block with concrete-enough wording (e.g., "5-stage timeline 가로 막대, 색상 5개"). PPTX export is NOT performed by this pipeline; the user reads the cheatsheet markdown and creates slides manually in PowerPoint with their lab template.
• Present material inventory to the user briefly and auto-proceed.

Task

$ARGUMENTS