Academic Orchestrator

Multi-phase orchestrator for academic research and writing. Does NOT do research or writing itself — delegates each phase to the appropriate specialist skill, enforces quality gates, and persists all intermediate artifacts to disk.

Critical design decision: The full 8-phase pipeline exceeds a single Claude Code session's context budget. Each skill invocation alone costs 5-15K tokens to load, agent results can reach 30K tokens each, and the draft + fact-check both need the full manuscript in context. The pipeline is therefore split across 3 sessions bridged by files on disk, regardless of output format. Each session starts clean and loads only what it needs from previous outputs.

Context Budget (Measured from Test Run)

Session	Phases	Measured Token Range	Composition
1	1, 2, 2.4, 3	60-85K	Orchestrator (13K) + 3 search agents (30-50K) + writing skill (8K) + academic-writing (3K) + 2 deep-read agents* (0K main session, agents run bg) + enriched draft (6K) + discussion
2	4, [+5 if LaTeX]	15-35K	citation-management (8K) + bib validation + retraction check + [latex-paper-en (8K) + .tex]
3	6, 7, 8	40-65K	fact-check (11K) + adversarial + manuscript (5K) + 3 reviewers parallel + elements-of-style (2K)

*Phase 3.1b deep-read agents run in background with independent contexts. Their skill instructions are inline in the prompt, not loaded into the main session. Cost to main session: only the agent launch messages (~2K) + the enriched draft diff (+1-2K over v1). Net Session 1 increase: ~3K.

Rules to stay within budget:

• After each session: immediately run /compact to summarize before the next
• Never load a skill's SKILL.md unless actually invoking it in that session
• Agent result disposal: After writing agent results to files (Phase 2.2 and Phase 3.1b), do NOT keep raw agent output in working memory. Read only the merged/enriched files for subsequent phases
• Session 3: fact-check the manuscript section by section, don't load all at once
• Skill-as-Agent pattern: for tasks that would load a large skill (analyzing-research-papers, peer-review), use Agent tool with inline instructions instead of Skill tool. The skill's methodology goes into the agent prompt — the main session never loads the skill file
• Skill invocation costs (loaded in main session): orchestrator (13K), literature-review (4K), medical-imaging-review (8K), academic-writing (3K), citation-management (8K), fact-check (11K), latex-paper-en (8K), elements-of-style (2K)
• Skills NOT loaded in main session (used via Agent tool): analyzing-research-papers (4K would-be), peer-review (3K would-be)
• MCP servers added to ~/.claude.json only take effect after restarting Claude Code. If agents report a new MCP tool is unavailable, the user needs to restart. Agent prompts include PRIMARY/FALLBACK tool instructions so search quality degrades gracefully when MCP tools are not yet loaded

Architecture

Two paths diverge at Phase 1. RESEARCH-ONLY is single-session, lightweight. Full pipeline is 3-session, comprehensive.

RESEARCH-ONLY Path (Single Session — ~35-45K tokens)

RESEARCH-ONLY (single session)
  Phase 1  SCOPE    → Clarify + route to RESEARCH-ONLY
  GATE 1: Confirm research plan
  Phase 2  RESEARCH → Parallel multi-source (Agent × 3, bg)
                      → Write to files → Clear → Merge
                      → Citation chaining (S2 MCP)
  Phase R  OUTPUT   → Generate directly (no external skill)
                      → Single file: research-digest.md
                      → Topic-adaptive tables + key findings + sources
  GATE R: Review digest → refine or deliver
  Output: research-output/research-digest.md

Full Pipeline (3 Sessions)

SESSION 1 (research + draft) — ~55-80K tokens
  Phase 1  SCOPE    → Clarify + route
  GATE 1: Confirm research plan
  Phase 2  RESEARCH → Parallel multi-source (Agent × 3, bg)
                      → Write to files → Clear from memory → Merge
  Phase 3  DRAFT    → Academic writing
  GATE 2: Review draft
  Output: research-output/*.md
  END: /compact

SESSION 2 (citations + [format]) — ~15-35K tokens
  Load from files: phase3-draft.md, phase2-merged.md
  Phase 4  CITATIONS → .bib construction + DOI validation
  Phase 5  FORMAT    → [FULL: LaTeX conversion + diagnostics]
                       [Markdown-only: skipped — proceed to /compact]
  Output: manuscript.tex + references.bib, OR validated draft.md
  END: /compact

SESSION 3 (verify + review + final) — ~40-65K tokens
  Load from files: manuscript.tex (or phase3-draft.md), references.bib
  Phase 6  VERIFY    → Fact-check + adversarial verification
  GATE 3: Review verification results
  Phase 7  REVIEW    → 3 parallel peer reviewers → merge consensus
  GATE 4: Review consensus → decide on revisions
  Phase 8  FINAL     → Language polish + final output
  GATE 5: Final sign-off
  Output: Corrected manuscript, VERIFICATION_STATUS.md

Skill Dispatch Matrix

RESEARCH-ONLY Path (single session)

Phase	Skill to Invoke	Tool	Notes
1	(Agent directly) + literature-review	AskUserQuestion	Clarify scope; adopt PRISMA/PICO framework
2	deep-research, academic-researcher, [+ medical-imaging-review]	Agent (bg, parallel) + S2 MCP	Full parallel search + citation chaining
R	(Agent directly)	Write	Structured tables + key findings + annotated sources (templates in Phase R spec)

Full Pipeline (3 sessions)

Phase	Skill to Invoke	Tool	Notes
1	(Agent directly) + literature-review	AskUserQuestion	Clarify scope; adopt PRISMA/PICO framework
2	deep-research, academic-researcher, [+ medical-imaging-review]	Agent (bg, parallel) + S2 MCP	Full parallel search + citation chaining
3	3.1: domain skill → 3.1b: deep-read → 3.1c: code audit (opt) → 3.1d: figures+tables (opt) → 3.2: ∥ prose ∥ citations ∥ data licensing	Skill + MCP + Agent bg (3.2)	Serial enrich → parallel refine → merge
4	citation-management	Skill + WebFetch	.bib + retraction check + source quality annotation
5	latex-paper-en	Skill	Convert to .tex (FULL only)
6	fact-check	Skill	Verification + adversarial counter-evidence
7	peer-review (×3 parallel personas)	Agent (bg, parallel)	Methodologist + Domain Expert + Generalist
8	elements-of-style + citation-management + [latex-paper-en]	Skill	Language polish → citation validation → [LaTeX]

---

Phase 1: Scope & Route

Step 1.1: Extract Requirements

Ask the user (use AskUserQuestion):

• "What is the topic and research question?"
• "What type of output?" — Research digest (tables + summaries, no paper) / Survey paper / Systematic review / Research proposal / Course paper
• "Target venue?" (if paper) — Conference / Journal / arXiv preprint / Course submission
• "Depth level?" — Quick overview / Standard review / Exhaustive systematic review
• "Any specific sources, papers, or angles to include/exclude?"

If the user selects "Research digest": route to RESEARCH-ONLY strategy. This is a single-session, lightweight path that produces structured tables, annotated source lists, and a key-findings summary — no draft writing, no LaTeX, no peer review.

Step 1.2: Domain Routing

Output type is "Research digest"?
  ├─ YES → Strategy: RESEARCH-ONLY
  │         Phase 2: deep-research + academic-researcher [+ medical-imaging-review]
  │         Phase R: research-synthesis skill → tables + summary + sources
  │         Single session, ~35-45K tokens
  │
  └─ NO → Topic is medical imaging AI (CT, MRI, X-ray, ultrasound, pathology)?
            ├─ YES → Strategy: MEDICAL
            │         Phase 2: deep-research + academic-researcher + medical-imaging-review
            │         Phase 3: medical-imaging-review (primary writer)
            │         Full 8-phase pipeline (3 sessions)
            │
            └─ NO → Is the topic academic/scholarly?
                      ├─ YES → Strategy: ACADEMIC
                      │         Phase 2: deep-research + academic-researcher
                      │         Phase 3: academic-researcher (primary writer)
                      │         Full 8-phase pipeline (3 sessions)
                      │
                      └─ NO → Strategy: GENERAL
                                Phase 2: deep-research only
                                Phase 3: academic-researcher
                                Full 8-phase pipeline (3 sessions)

Step 1.3: Output the Research Plan

Write research-output/phase1-plan.md:

# Research Plan: [Topic]
- **Strategy**: [MEDICAL/ACADEMIC/GENERAL]
- **Research questions**: [3-5 sub-questions]
- **Skills to invoke**: [list]
- **Output type**: [survey/systematic/proposal/paper]

GATE 1: Present the plan to the user. Do NOT proceed to Phase 2 until the user confirms.

---

Phase 2: Multi-Source Parallel Research

Why Parallel (Not Sequential)

Each skill uses a disjoint source pool — they search different corners of the internet:

• deep-research: Firecrawl + Exa → general web, news, industry reports, blogs
• academic-researcher: Scholarly sources → peer-reviewed papers, structured analysis, citations
• medical-imaging-review: arXiv + PubMed + Zotero → domain-specific literature (MEDICAL only)

Running them sequentially is not just slower — it introduces bias. If deep-research runs first and finds X, academic-researcher may anchor on X and miss Y. Running them blind to each other, then cross-validating, catches more and overweights less.

Cost-benefit: 2-3 parallel agents instead of 1, but wall-clock time is ~the slowest single agent (60-90s), not the sum. Coverage gain is substantial — our test showed Agent 1 found physics-inspired attention mechanisms and frequency-domain approaches that Agent 2 missed, while Agent 2 found a specific ICLR paper and implementation details that Agent 1 missed. Only overlap: sparse attention / Focus trend. Combined coverage was ~3x either alone.

Step 2.1: Launch Agents in a Single Message

Critical: Launch ALL agents by putting multiple Agent tool calls in ONE message. This is what makes them truly concurrent — each gets its own context window and runs independently. Use run_in_background: true so the main session is not blocked.

For each sub-question from Phase 1, launch:

Message to user: "Launching N parallel research agents for: [sub-question]..."

Agent tool call 1 (run_in_background: true):
  subagent_type: "general-purpose"
  description: "deep-research: [sub-question]"
  prompt: |
    You are doing multi-source web research. Search for: "[sub-question]"
    - Use WebSearch and/or firecrawl_search for broad coverage
    - Focus on: latest developments, news, industry reports, blog posts, non-academic sources
    - Find 5-10 key sources
    - Return your findings AS TEXT in your response. Structure them as:
      ## [Sub-question] — Web Perspective
      ### Key Findings
      1. [Finding] — Source: [title](url)
      2. ...
      ### Sources
      [numbered list with URLs]
    - Do NOT try to write files. Just return the text in your response.

Agent tool call 2 (run_in_background: true):
  subagent_type: "general-purpose"
  description: "academic-researcher: [sub-question]"
  prompt: |
    You are doing academic literature research. Search for: "[sub-question]"
    - REQUIRED: First call mcp__paper-search__search_arxiv(query="[sub-question]", maxResults=10) AND mcp__paper-search__search_pubmed(query="[sub-question]", max_results=10). EFFICIENCY: Issue ALL independent MCP calls in ONE message (batch them) to cut wall-clock time by 3-4x. CRITICAL: Pass a real keyword query — empty query returns daily-new-papers, not matches.
    - Also call mcp__paper-search__search_google_scholar(query="[sub-question]") for comprehensive coverage. For bio/medical topics, call mcp__paper-search__search_biorxiv + mcp__paper-search__search_medrxiv with the same query
    - Only if ALL MCP calls fail or return empty: fall back to WebSearch + firecrawl_search + exa
    - Focus on: peer-reviewed papers, methodology, experimental results, citations
    - Find papers: 5-10 (Quick overview) / 10-15 (Standard) / 15-20 (Exhaustive). Scale to the depth level from Phase 1.
    - Return your findings AS TEXT in your response. Structure them as:
      ## [Sub-question] — Academic Perspective
      ### Key Papers
      1. [Paper title] ([Year]) — [1-line finding]. DOI/URL: [link]
      2. ...
      ### Methodological Themes
      [patterns across papers]
      ### Research Gaps
      [what's missing]
    - REPORT which tools you used: "[MCP: USED arxiv+pubmed]" or "[FALLBACK: reason]". Do NOT try to write files.

Agent tool call 3 (run_in_background: true) — MEDICAL strategy ONLY:
  subagent_type: "general-purpose"
  description: "medical-imaging: [sub-question]"
  prompt: |
    You are doing medical imaging literature research. Search for: "[sub-question]"
    - REQUIRED: First call mcp__paper-search__search_pubmed AND mcp__paper-search__search_google_scholar(query="[sub-question]")(query="[sub-question]", max_results=10) AND mcp__paper-search__search_medrxiv(query="[sub-question]", max_results=10) AND mcp__paper-search__search_biorxiv(query="[sub-question]", max_results=10). EFFICIENCY: Issue ALL independent MCP calls in ONE message (batch them) to cut wall-clock time by 3-4x. CRITICAL: Pass a real keyword query to each — empty query returns noise.
    - Only if ALL three MCP calls fail or return empty: fall back to firecrawl_search + WebSearch + exa
    - Focus on: clinical validation, Dice/HD95 metrics, public datasets used
    - Find papers: 5-10 (Quick overview) / 10-15 (Standard) / 15-20 (Exhaustive). Scale to the depth level from Phase 1.
    - Return your findings AS TEXT in your response. Structure them as:
      ## [Sub-question] — Medical Imaging Perspective
      ### Key Papers
      1. [Paper title] ([Year]) — Method: [method], Dice: [score], Dataset: [dataset]. URL: [link]
      ...
    - REPORT which tools you used: "[MCP: USED pubmed+medrxiv+biorxiv]" or "[FALLBACK: reason]". Do NOT try to write files.

Step 2.2: Collect Results, Write to Disk, Clear from Memory

When ALL agents complete, for each agent:

1. Extract the full findings from the completion notification text
2. Write to its file immediately — do NOT truncate or summarize: research-output/phase2-deep-research.md, research-output/phase2-academic-researcher.md, research-output/phase2-medical-imaging.md (if MEDICAL)
3. After writing: clear the raw agent output from working memory. The files on disk are the authoritative record

Why the main session writes files, not agents: Background agents may lack Write/Bash permissions. Main session persists them.

Why dispose of raw output: Three agent results can total 30-50K tokens. Keeping them in working memory alongside the orchestrator, writing skill, and draft would exhaust the context budget before Phase 3 even begins. The files are on disk — subsequent phases read only the merged synthesis.

Step 2.3: Merge & Cross-Validate (from files only)

Read phase2-deep-research.md, phase2-academic-researcher.md, and phase2-medical-imaging.md from disk. Do NOT use the raw agent completion text still in conversation memory — read the files. Then produce research-output/phase2-merged.md:

# Merged Research Notes: [Topic]

## Agreements (found by 2+ skills → HIGH confidence)
- [Claim] — Sources: [d-r ref], [a-r ref]

## Unique Findings (found by 1 skill only)
### From deep-research
- [Finding] — Source: [ref]

### From academic-researcher
- [Finding] — Source: [ref]

## Contradictions (flagged for investigation)
- [Skill A says X] vs [Skill B says Y] — Resolution: [which is more credible and why]

## Source Inventory
| # | Title | Type | Source Skill | URL/DOI |
|---|-------|------|-------------|---------|

Step 2.4: Citation Chaining (Discovery Beyond Keywords)

Goal: Keyword search misses papers that don't use the same terms. Citation chaining finds them through the citation graph.

1. From the merged source inventory, identify the top-5 most impactful papers (highest citation counts, seminal works, recent highly-cited surveys)
2. Use the semantic-scholar MCP tools (mcp__semantic-scholar__papers-search-basic...) to:
   • Forward search: Find papers that cite these top-5 papers → discover latest developments building on seminal work
   • Backward search: Extract the reference lists of these top-5 papers → find foundational works that keyword search may have missed
3. For any newly discovered papers that are highly relevant, add them to the source inventory in phase2-merged.md
4. Append a "Citation Graph Discoveries" section to the merged file listing 3-5 newly found papers

Why this matters: Our test run found 26 sources via keyword search alone. Citation chaining on nnU-Net and TotalSegmentator would discover papers that build on these methods but use different terminology — filling a known gap in keyword-only discovery.

Quality Rules for Phase 2

• Every claim must have a source attached
• Contradictions must be explicitly flagged, not silently dropped
• If only one skill found a claim, mark it as "single-source — lower confidence"
• Evidence strength ladder: Annotate each source in the merged inventory with its evidence level:
  • [A] Peer-reviewed journal article / top conference (NeurIPS, CVPR, MICCAI)
  • [B] Peer-reviewed conference workshop / lower-tier journal
  • [C] Preprint (arXiv, bioRxiv, etc.) — not yet peer-reviewed
  • [D] Grey literature (industry report, blog post, corporate website)
  • When two sources disagree, higher evidence level prevails

---

Phase R: Research Digest Output (RESEARCH-ONLY strategy)

Phase R replaces Phases 3-8 for RESEARCH-ONLY strategy. The user wants a single, scannable research digest — not a paper.

Step R.1: Choose Output Format

Read the user's research question and the merged findings. Choose ONE primary format that best answers their question. Do NOT default to tables — pick the format that communicates most effectively.

Question Type	Best Format	Example
"Compare A vs B"	Comparison table or side-by-side pros/cons	"对比 nnU-Net 和 SAM 在器官分割上的表现"
"What methods exist?"	Categorized list with strengths/weaknesses	"全身器官分割有哪些方法"
"What's the consensus?"	Agreement matrix or "What we know / What's contested / What's missing"	"领域共识是什么"
"How did we get here?"	Timeline or milestone list	"这个领域的发展历程"
"What should I do?"	Recommendations list with rationale and confidence	"我该用什么方法"
"What data is available?"	Dataset table	"有哪些公开数据集"
"What's the latest?"	Reverse-chronological highlights or trend summary	"最新进展"
"Mixed / broad topic"	Executive summary + key findings + whatever formats fit each subsection	"调研全身器官分割"

Step R.2: Generate the Digest

Generate research-output/research-digest.md. All content in ONE file. The skeleton adapts to the chosen format, but always includes:

# Research Digest: [Topic]
*Date | Sources: [N] | Strategy: RESEARCH-ONLY | Confidence: [HIGH/MEDIUM/LOW]*

## What We Know *(required)*
[3-5 sentence executive summary. Answers the user's question directly — don't bury the lead.]

## [Body — format chosen in Step R.1] *(required)*
*This section's structure matches the question type. Examples:*
- Comparison table (methods/datasets/tools side-by-side)
- Categorized list (by approach, by evidence strength, by chronology)
- Agreement matrix (consensus vs contested vs unknown)
- Timeline (year, milestone, significance, source)
- Recommendations (action, rationale, confidence, caveats)
*Tables are ONE option among many. Use the format that communicates best.*

## What's Contested / What's Missing *(required)*
[3-5 gaps, contradictions, or open questions. Every research digest should tell the user what we DON'T know.]

## Cross-Validation Summary *(required)*
| Claim | deep-research | academic-researcher | [medical] | Strength |
|-------|--------------|--------------------|-----------|----------|
| [Claim] | ✅ | ✅ | ✅ | HIGH |
| [Claim] | ✅ | ❌ | — | Single |

## Annotated Sources *(required)*
| # | Title (Author, Year) | Type [A/B/C/D] | Found By | 1-Line Relevance |
|---|---------------------|----------------|----------|-----------------|

*Sources sorted by evidence level then recency. Full URLs/DOIs inline.*

## Raw Research Files *(required — for traceability)*
- `research-output/phase2-deep-research.md` — Full web perspective notes
- `research-output/phase2-academic-researcher.md` — Full academic perspective notes
- `research-output/phase2-medical-imaging.md` — Full medical imaging notes (if MEDICAL strategy)
- `research-output/phase2-merged.md` — Cross-validated synthesis with evidence levels

*If any claim in this digest needs verification or more detail, the raw files above contain the original agent output.*

Design Principles

• Format follows question: don't force methods into a table or trends into a list. Pick the format that answers the user's actual question
• Lead with the answer: executive summary states the conclusion first, not the process
• What we DON'T know is as important as what we do: every digest must include gaps and contested claims
• Evidence levels everywhere: every claim/source tagged [A/B/C/D]. Every finding tagged with consensus strength
• Single file, scannable in 60 seconds: no separate files, no redundant content across sections
• Back-linked to raw research: the digest footer must include a "Raw Research Files" section linking to phase2-deep-research.md, phase2-academic-researcher.md, etc. for traceability

Step R.3: Language Polish (Prose Sections Only)

After generating the digest, invoke writing-clearly-and-concisely skill (Elements of Style) for a quick language pass on the prose sections (Executive Summary, Key Findings, Gaps). This is lightweight (~2K tokens) and catches passive voice, wordiness, and unclear phrasing. Do NOT apply to tables or source list — those are structural, not prose.

---

Phase 3: Multi-Pass Draft Writing (FULL pipeline only — not RESEARCH-ONLY)

Writing is inherently sequential at the structural level (you can't parallelize the act of composing a single narrative), but refinement passes on different dimensions CAN run in parallel. Phase 3 uses a serial draft + parallel refinement + merge pattern.

Step 3.1: Structural Draft (Single Writer)

Invoke the primary writing skill via the Skill tool:

• MEDICAL strategy → medical-imaging-review
• ACADEMIC or GENERAL strategy → academic-researcher

Provide as context: research-output/phase2-merged.md and the skill's own template. Output to research-output/phase3-draft-v1.md (v1 = structural draft only).

Step 3.1b: Deep-Read Key Sources (Agent-Parallel, Enrich from Primary Sources)

The structural draft (v1) was written from Phase 2 summaries — second-hand information. This step deep-reads the most important papers to fill in exact numbers, method specifics, and author-stated caveats that summaries may have omitted.

Context management: Deep reading loads a 4K skill if done in the main session. Instead, use the same Agent-parallel pattern as Phase 2 — launch background agents with inline instructions. The main session never loads analyzing-research-papers. Cost to main session: ~2K for agent launch + ~1-2K for the enriched draft diff.

How many to deep-read? ~30-40% of sources, minimum 8, maximum 15. Select by: [A][B] evidence level only → sort by citation count (S2 MCP) → ensure each sub-topic has ≥1 paper represented.

Launch 2 agents in parallel (split the paper list evenly). Each agent gets this prompt:

You are deep-reading academic papers to extract specific details for a survey draft.
You have access to: Paper Search MCP, S2 MCP, and WebFetch.

For each paper in your list:
1. Paper Search MCP get_paper_details(DOI) → get metadata + abstract
2. S2 MCP → get citation count + TLDR → confirm this paper is high-impact
3. WebFetch the paper URL → extract these 4 things from the full text:
   a) EXACT METRICS: Dice scores, HD95, p-values, sample sizes — specific numbers
   b) METHOD DETAILS: architecture, training protocol, data splits, augmentations
   c) AUTHOR-STATED LIMITATIONS: their own caveats, in their own words
   d) 1-2 CUTABLE QUOTES: with section reference

Return your findings AS TEXT. Structure per paper:
### [Paper Title] (Author, Year)
- **Metrics**: [exact numbers]
- **Methods**: [architecture specifics, training details]
- **Limitations**: [author-stated caveats]
- **Quotes**: "[exact quote]" (Section X)

Do NOT evaluate the paper's quality. Do NOT write prose. Just extract facts.
Do NOT try to write files — return text inline.

Paper list: [paste DOIs/URLs for 4-7 papers]

When both agents complete, collect their findings, write to research-output/phase3-deep-reads.md, then clear raw agent output from working memory.

Finally, apply the findings to v1 → output phase3-draft-v1-enriched.md. Fill in missing numbers, add method specifics, insert author-stated caveats and quotes. This is a factual patch — do NOT restructure or rewrite.

Step 3.2: Parallel Refinement (3 Agents, Background)

Launch 3 Agent tasks IN A SINGLE MESSAGE with run_in_background: true. Each refines the draft on a different, independent dimension:

Agent A — Prose Quality (academic-writing perspective):
  prompt: |
    Read the draft below. Focus ONLY on language quality:
    - Remove hedging soup (every paragraph ≤2 hedging words: potentially, may, might, could, arguably, perhaps)
    - Replace formulaic transitions (Furthermore, Moreover, Additionally, It is important to note that) with logical flow
    - Strengthen specificity: replace abstract claims with named studies, sample sizes, years
    - Inject authorial voice where the discipline permits
    - Do NOT change structure, facts, citations, or argument. Only language.
    Return the refined prose inline as text.

Agent B — Citation Completeness (literature-review perspective):
  prompt: |
    Read the draft below. Focus ONLY on citation coverage:
    - Are all factual claims backed by a citation?
    - Are there missing seminal works that should be cited?
    - Are any citations attributed to the wrong paper?
    - Check against the merged research notes (phase2-merged.md) for cited-but-not-in-sources
    Return a checklist: [MISSING] for missing citations, [WRONG] for misattributed, [OK] for correct.
    Do NOT rewrite prose. Just return the citation audit.

Agent C — Data & Licensing Audit (medical-imaging domain focus):
  prompt: |
    Read the draft below. Focus ONLY on dataset provenance and licensing:
    For EVERY dataset mentioned in the draft, extract and verify:
    - **Dataset name and source repository** (e.g., TCIA collection ID, Zenodo DOI, GitHub release)
    - **Coverage / FOV**: anatomical regions covered, organs included, any notable exclusions
    - **License type**: CC BY 3.0, CC BY 4.0, CC BY-NC (non-commercial only), custom/proprietary, or unspecified
    - **Commercial use**: explicitly permitted, restricted, or unclear from the license
    - **Key caveats**: pediatric patients excluded? pathology cases excluded? single-vendor only?
    
    Return a structured audit. For each dataset:
    ```
    ### [Dataset Name]
    - **Source**: [URL / DOI / TCIA ID]
    - **Coverage**: [FOV description]
    - **License**: [CC BY 3.0 / CC BY 4.0 / CC BY-NC / Proprietary / Unspecified]
    - **Commercial use**: [Permitted / Restricted / Unclear]
    - **Caveats**: [any limitations on scope, population, or vendor]
    ```
    
    Also flag any dataset mentioned WITHOUT an explicit citation or source link → [UNVERIFIED-DATASET].
    Do NOT rewrite prose. Just return the data audit.

Step 3.3: Merge Refinements

When all 3 agents complete:

1. Apply Agent A's language refinements to v1-enriched → phase3-draft-v2.md
2. Apply Agent B's citation fixes to v2 → phase3-draft-v3.md
3. Apply Agent C's data notes to v3 → phase3-draft.md (final):
   • Add license/caveat annotations to dataset descriptions
   • Flag any [UNVERIFIED-DATASET] items as caveats in the text
4. Record audits: research-output/phase3-citation-audit.md (Agent B), research-output/phase3-data-licensing-audit.md (Agent C)
5. Clear raw agent output from working memory after writing files

Why parallel works here: Prose, citations, and data licensing are three orthogonal dimensions — one changes words, one checks references, one verifies dataset provenance. They don't conflict. Running them sequentially would take 3x the wall-clock time with zero quality gain.

Draft Quality Minimums

• Topic sentences with clear claims
• Every factual claim has at least an inline citation marker
• Method descriptions include limitations
• Comparison table for each major section (if ≥ 3 items to compare)
• Survey Methodology subsection: State search period, databases/sources used, keywords, inclusion/exclusion criteria, and the multi-source cross-validation approach
• Citation audit passed: All [MISSING] and [WRONG] items from Step 3.2 Agent B resolved
• Data & licensing audit passed: All datasets have verified sources and license annotations; [UNVERIFIED-DATASET] items flagged as caveats in the text

Step 3.1c: Code Repository Audit (OPTIONAL — only if user requests reproduction/code details)

Trigger: Run when the user shows ANY intent to go beyond reading the paper: code access, implementation details, reproduction, testing, training requirements, GPU specs, pretrained weights, licensing. Example: "I want to try this", "can I run this", "what GPU", "is the code available", "how do I train this", "any pretrained weights".

No dedicated MCP or skill exists for this purpose. The ecosystem has repo-scan (1.7K installs) but it scans LOCAL codebases (C++/Java/iOS), not remote paper repos. Use built-in tools:

Need	Tool	How
Find paper's GitHub repo	gh search repos "[paper title]" --limit 5	Bash
	or WebSearch "[paper title] github"	WebSearch
Read README	WebFetch https://raw.githubusercontent.com/.../README.md	WebFetch
Read training config	WebFetch configs/*.yaml or *.json (in repo)	WebFetch
Read dependencies	WebFetch requirements.txt or environment.yml	WebFetch
Check for pretrained weights	WebSearch [model name] pretrained weights download	WebSearch
Check inference demo	WebFetch repo tree → look for demo.py, inference.py, predict.py, Colab link	WebFetch

REPO VERIFICATION — do this first, before auditing:

1. gh search repos "[paper title]" --limit 5 → get candidate URLs
2. For each candidate, WebFetch its README. Verify AT LEAST 2 of: README mentions paper title/DOI, repo owner matches paper first author or lab, README describes the paper's method, repo has ≥10 stars or recent commits
3. No candidate passes → mark "[NOT FOUND — no verified public repo]"
4. Candidate passes → confirm with second file check (requirements.txt or setup.py exists)

8-POINT AUDIT — only on verified repos:

□ Official repo URL: [verified github.com/...] or "[NOT FOUND — no verified public repo]"
□ Pretrained weights: [URL] or "[NOT FOUND]" or "[in repo — download script]"
□ GPU requirement: [X GB VRAM / "not stated"] — search README + configs for "GPU", "memory", "batch"
□ Training specifics: [unique loss / custom scheduler / gradient clip value / mixed precision] — from train config
□ Inference demo: [filename] or "[NOT FOUND — no demo script]"
□ Dependency pinning: [pinned / unpinned] — check requirements.txt for == vs >=
□ Data preprocessing: [script exists / documented only / not provided]
□ License: [MIT / Apache / CC BY-NC / custom / not stated]

Save to research-output/phase3-code-audit.md. Add key findings (GPU requirements, license restrictions, pretrained weight availability) as implementation notes in the draft.

Why top-3 only: Auditing 3 repos involves 10-15 WebFetch calls. More than that adds significant context pressure and wall-clock time with diminishing returns — the top papers' repos cover the core implementation patterns.

Step 3.1d: Figure & Table Generation (OPTIONAL — for FULL pipeline papers)

Trigger: When the draft needs comparison charts, performance tables, or method diagrams. Not for RESEARCH-ONLY (digest already has tables). Agent judges: if the draft has ≥3 comparable methods or ≥5 comparable metrics, suggest figure/table generation to the user.

Available skills (both installed):

• figure-generation (137 installs): 10 chart types (bar, line, scatter, heatmap, radar, violin, training-curve, ablation, tsne, attention). Self-healing pipeline: generates Python script → executes → captures errors → retries up to 4×. Outputs PNG (300 DPI) + PDF (vector) + LaTeX include code. Colorblind-friendly palette.
• table-generation (same repo): Formatted publication-quality tables from research data. Converts JSON/CSV to LaTeX booktabs-style with bold best results.

Process:

1. Identify 2-4 figures/tables needed (comparison chart, architecture timeline, performance table)
2. For each: describe what to visualize + provide data from phase3-deep-reads.md
3. Invoke figure-generation or table-generation skill
4. Save to figures/ directory. Reference in draft with LaTeX \includegraphics or Markdown image links

Quality: ≥300 DPI, colorblind-friendly, labels ≥8pt, no matplotlib default titles.

GATE 2: Present the draft summary to the user. "Draft complete — [N] words, [M] sources. Citation audit: [N] missing, [M] misattributed (all fixed). Review before verification?" Do NOT proceed until the user confirms.

END OF SESSION 1

Before closing this session:

1. Verify these files exist on disk: research-output/phase1-plan.md, research-output/phase2-merged.md, research-output/phase3-draft.md
2. Tell the user: "Session 1 complete. All research and draft are saved to research-output/. Run /academic-orchestrator again and say 'continue from Phase 4' or start Session 2 with: claude -r "session-1-name" "continue the academic orchestrator from Phase 4""
3. Run /compact to summarize the conversation before starting Session 2

---

Session 2: Citations & Formatting

Starting Session 2

When the user returns for Session 2, you are in a fresh context. Immediately:

1. Read research-output/phase3-draft.md and research-output/phase2-merged.md (these contain everything you need — do NOT re-read other Phase 2 files unless needed)
2. Do NOT re-run any Phase 1 or 2 work
3. Confirm to user: "Loaded draft ([N] words, [M] sources). Ready for Phase 4."

---

Phase 4: Citation Management + Source Quality

Invoke citation-management skill via the Skill tool.

Input: The inline citation markers and source inventory from research-output/phase3-draft.md.

Tasks:

1. For each source, resolve DOI/PMID/arXiv ID → full metadata
2. Generate references.bib with all entries validated
3. Check for: missing required fields, duplicate entries, broken DOIs
4. Format BibTeX consistently (standardize field order, author names, capitalization)

Additional quality checks (performed by the main agent, not citation-management):
5. Retraction check: For the top-10 most-cited or most-critical sources, WebFetch https://pubmed.ncbi.nlm.nih.gov/?term=[paper title] retraction or similar. Flag any retracted papers in the citation report
6. Preprint → published upgrade: For arXiv preprints, check if a peer-reviewed journal version exists (via Semantic Scholar MCP or CrossRef). If yes → use the published version
7. Source quality annotation: Extend the evidence ladder from Phase 2.3 into the citation report. Mark each reference with its evidence level [A/B/C/D]

Output:

• references.bib — cleaned, validated BibTeX file
• research-output/phase4-citation-report.md — issues found, retraction status, evidence levels

---

Phase 5: LaTeX Formatting with Self-Healing Compilation

Step 5.1: Produce .tex File

If the draft is in Markdown:

1. Convert to LaTeX using the target venue's document class or a generic article class
2. Write to manuscript.tex
3. Ensure all \cite{} commands reference keys in references.bib

Step 5.2: Self-Healing Compilation Loop

LaTeX code is brittle — unescaped special characters, missing \usepackage, or BibTeX mismatches cause fatal errors. A single diagnostic pass catches some issues, but the only proof of compilability is a successful pdflatex run. This step implements a compile → diagnose → fix → recompile loop with max_retries = 3.

Loop algorithm:

attempt = 0
while attempt < 3:
    1. Run compilation in Bash:
       pdflatex -interaction=nonstopmode manuscript.tex 2>&1 | tee /tmp/latex-output.log
       bibtex manuscript 2>&1 >> /tmp/latex-output.log
       pdflatex -interaction=nonstopmode manuscript.tex 2>&1 >> /tmp/latex-output.log
       pdflatex -interaction=nonstopmode manuscript.tex 2>&1 >> /tmp/latex-output.log

    2. Check exit code and log:
       if pdflatex returns 0 AND "Fatal error" NOT in log:
           → SUCCESS. Exit loop.
       else:
           attempt += 1
           if attempt == 3:
               → FAIL. Write the last error log to research-output/phase5-compile-errors.log.
                 Report to user: "LaTeX compilation failed after 3 attempts. 
                 Error log saved. Continue with Phase 6 on the last known-good draft?"
                 Do NOT block the pipeline — proceed to Phase 6 with the .md draft as fallback.

    3. Extract the FIRST fatal error from /tmp/latex-output.log:
       - grep for lines containing "! " (LaTeX error marker) or "Fatal error"
       - Extract the 5 lines before and after the error for context
       - Identify error type: undefined control sequence / missing package / 
         bad character escape / BibTeX mismatch / missing file

    4. Feed error to Agent for targeted fix:
       Launch a SINGLE Agent with the error context:
       "The LaTeX compilation failed with this error: [paste error block]. 
        Fix ONLY this specific error in manuscript.tex. Do NOT rewrite the document.
        Common fixes: escape underscores in non-math contexts, add missing \\usepackage{X}, 
        fix unbalanced braces, replace Unicode chars with LaTeX equivalents."

    5. Apply the Agent's fix to manuscript.tex
    6. Retry from step 1

Why this works: LaTeX errors are usually atomic — one bad \usepackage, one unescaped _, one mismatched brace. Fixing the first error often resolves cascading errors downstream. The loop targets one error per iteration, which is more reliable than trying to fix everything at once.

Step 5.3: Final Diagnostic Pass

Once compilation succeeds, invoke latex-paper-en skill via the Skill tool for final polish:

• Structure check (abstract/conclusion alignment)
• Language polish (grammar, academic tone)
• Venue formatting compliance (if target specified)

Output: manuscript.pdf (compiled) + manuscript.tex (compilable source).

Session 2 Checkpoint: Report to user (FULL strategy only). "LaTeX compilation: [SUCCESS after N attempts / FAILED — see phase5-compile-errors.log]. PDF generated: [yes/no]." For Markdown-only: skip Phase 5 entirely, proceed directly to /compact.

END OF SESSION 2

Before closing this session:

1. Verify these files exist on disk: manuscript.tex, references.bib, research-output/phase4-citation-report.md
2. Tell the user: "Session 2 complete. manuscript.tex and references.bib are ready. Run /academic-orchestrator again and say 'continue from Phase 6' to start verification."
3. Run /compact

---

Session 3: Verification & Review

Starting Session 3

When the user returns for Session 3, you are in a fresh context. Immediately:

1. Read manuscript.tex (this is your primary working file)
2. Read references.bib (needed for citation verification)
3. Do NOT re-read Phase 2/3 research files unless a specific claim needs source re-checking
4. Confirm to user: "Loaded manuscript ([N] words, [M] references). Ready for Phase 6 verification."

Memory discipline for Session 3: The fact-check skill requires the full manuscript in context. To avoid overflow:

• Read manuscript.tex section by section during verification, not all at once
• After each section is verified, write corrections immediately
• Use references.bib only for citation lookups — don't hold it all in working memory

---

Phase 6: Fact-Check (Independent Pass)

Critical constraint: This MUST be a separate pass after writing is complete. Do NOT combine with Phase 3 or Phase 5.

Invoke fact-check skill via the Skill tool.

Provide manuscript.tex (or research-output/phase3-draft.md if LaTeX not yet generated).

Output to research-output/phase6-factcheck.md:

• Every verifiable claim extracted and categorized (Verifiable-Hard / Verifiable-Soft / Attribution / Inference)
• Each claim checked against its source
• Confidence: Confirmed / Partially Supported / Not Found / Contradicted
• Overall reliability: High / Medium / Low / Unreliable

Post-Verification Actions

• CONFIRMED claims → keep, add source citation
• PARTIALLY SUPPORTED → narrow the claim to match the source
• NOT FOUND → mark as unverified or remove
• CONTRADICTED → remove or correct immediately

Apply all corrections to manuscript.tex. Record changes in research-output/phase6-corrections.md.

Adversarial Verification (Counter-Evidence Search)

Fact-check verifies "is this claim supported by its cited source?" — but does NOT ask "is there evidence AGAINST this claim?" Adversarial verification closes this gap.

Process:

1. Take the top 3-5 HIGH-confidence claims from the fact-check report
2. For each claim, run a targeted search for counter-evidence: "[claim keywords] controversy criticism limitation rebuttal"
3. If counter-evidence is found → downgrade confidence to Medium, add caveat to the manuscript
4. If no counter-evidence → confidence confirmed as HIGH

Output: Append an "Adversarial Verification" section to research-output/phase6-factcheck.md with the results. This step takes ~5 minutes and catches the most dangerous type of error — consensus claims that the field has moved past.

GATE 3: Present fact-check summary + adversarial verification to the user.

Show: overall confidence level, number of claims confirmed/unverified/contradicted, and any corrections applied. Ask: "Shall I proceed to peer review with these corrections, or would you like to review the fact-check details first?" Do NOT proceed to Phase 7 until the user confirms.

---

Phase 7: Multi-Reviewer Peer Review

Three independent reviewers evaluate the manuscript in parallel — mirroring real academic peer review where 2-3 reviewers catch different issues and consensus strengthens signals. Same pattern as Phase 2: launch agents in ONE message, they run concurrently, then merge.

Step 7.1: Launch Reviewers in Parallel

Launch 3 Agent tasks IN A SINGLE MESSAGE with run_in_background: true. Each gets a distinct reviewer persona:

Agent tool call 1 (run_in_background: true):
  subagent_type: "general-purpose"
  description: "Reviewer A — Methodologist"
  prompt: |
    You are Reviewer A — a PhD-level methodological specialist. Read the manuscript below and evaluate it using the peer-review framework: dimensional scoring (1-5) on Novelty, Rigor, Impact, Clarity. Your PRIMARY focus is METHODOLOGICAL SOUNDNESS:
    - Are the methods correctly described and appropriate?
    - Are statistical claims properly supported?
    - Are there gaps in experimental validation?
    - Are the conclusions warranted by the evidence presented?
    Return your review AS TEXT. Structure:
    ## Reviewer A — Methodologist
    **Recommendation**: [Accept/Minor/Major/Reject]
    **Scores**: Novelty:[S] Rigor:[S] Impact:[S] Clarity:[S]
    **Major Issues**: [numbered list with specific evidence from the text]
    **Minor Issues**: [numbered list]
    Do NOT try to write files. Return text inline.
    
    Manuscript:
    [paste manuscript.tex content]
    Fact-check report: [paste phase6-factcheck.md summary]

Agent tool call 2 (run_in_background: true):
  subagent_type: "general-purpose"
  description: "Reviewer B — Domain Expert"
  prompt: |
    You are Reviewer B — a senior domain expert in the paper's specific field. Read the manuscript below. Your PRIMARY focus is DOMAIN ACCURACY AND COVERAGE:
    - Does the paper accurately represent the state of the field?
    - Are there missing seminal works or important recent papers?
    - Are the claims consistent with domain knowledge?
    - Is the clinical/practical framing accurate and useful?
    Return your review AS TEXT. Structure:
    ## Reviewer B — Domain Expert
    **Recommendation**: [Accept/Minor/Major/Reject]
    **Scores**: Novelty:[S] Rigor:[S] Impact:[S] Clarity:[S]
    **Major Issues**: [numbered list]
    **Missing Literature**: [papers the manuscript should cite but doesn't]
    **Minor Issues**: [numbered list]
    Do NOT try to write files. Return text inline.
    
    Manuscript:
    [paste manuscript.tex content]
    Fact-check report: [paste phase6-factcheck.md summary]

Agent tool call 3 (run_in_background: true):
  subagent_type: "general-purpose"
  description: "Reviewer C — Generalist / Editor"
  prompt: |
    You are Reviewer C — an experienced journal editor with a generalist perspective. Read the manuscript below. Your PRIMARY focus is CLARITY, STRUCTURE, AND ACCESSIBILITY:
    - Is the argument flow logical and easy to follow?
    - Is the writing clear and well-structured?
    - Would a non-specialist reader understand the contribution?
    - Are figures/tables well-designed and informative?
    - Is the abstract accurate and compelling?
    Return your review AS TEXT. Structure:
    ## Reviewer C — Generalist / Editor
    **Recommendation**: [Accept/Minor/Major/Reject]
    **Scores**: Novelty:[S] Rigor:[S] Impact:[S] Clarity:[S]
    **Major Issues**: [numbered list]
    **Structural/Clarity Issues**: [specific examples from the text]
    **Minor Issues**: [numbered list]
    Do NOT try to write files. Return text inline.
    
    Manuscript:
    [paste manuscript.tex content]
    Fact-check report: [paste phase6-factcheck.md summary]

Step 7.2: Merge Reviews (from files, clear raw results)

When ALL 3 reviewers complete, for each reviewer:

1. Extract the full review from the completion notification
2. Write to its file immediately: research-output/phase7-reviewer-a.md, phase7-reviewer-b.md, phase7-reviewer-c.md
3. After writing all 3 files: clear raw review text from working memory. The files on disk are the authoritative record

Then read the 3 files from disk and produce the merged report research-output/phase7-peerreview-merged.md:

# Peer Review — Consolidated Report

## Reviewer Recommendations
| Reviewer | Role | Recommendation | N | R | I | C |
|----------|------|------------------|---|---|---|---|
| A | Methodologist | [Verdict] | | | | |
| B | Domain Expert | [Verdict] | | | | |
| C | Generalist | [Verdict] | | | | |

## Consensus Issues (found by 2+ reviewers → MUST FIX)
1. [Issue] — Reviewers A, B

## Individual Issues (found by 1 reviewer → SHOULD FIX)
### From Reviewer A (Methodologist)
- [Issue]

### From Reviewer B (Domain Expert)
- [Issue]

### From Reviewer C (Generalist)
- [Issue]

## Missing Literature (from Reviewer B)
- [Paper title] — relevance: [why it should be cited]

## Final Recommendation
Weighted by reviewer role. If 2+ reviewers agree on the verdict → that verdict. If all 3 disagree → flag for user decision.

GATE 4: Present merged review to the user.

Show: consensus recommendation, score matrix, consensus issues vs individual issues. Ask: "Accept all consensus fixes? Override any individual reviewer suggestions?" Apply user's decisions. If Major Revision from 2+ reviewers: offer to loop back to Phase 3. If Reject from 2+: flag with specific reasoning.

Post-Review Revision

Address all consensus issues (must-fix) per user's decision at Gate 4. Individual reviewer suggestions are at the user's discretion. Apply revisions to the manuscript.

---

Phase 8: Final Output

Step 8.1: Final Validation

Invoke citation-management skill to run final validation on references.bib.

Step 8.2: Language Polish

Invoke writing-clearly-and-concisely skill (Elements of Style) for final language polish. This is a purely stylistic pass — it improves clarity, concision, and professionalism WITHOUT changing factual content or argument structure. If LaTeX output, invoke latex-paper-en skill after language polish for final formatting consistency check.

Step 8.3: Final LaTeX Pass (FULL strategy only)

Invoke latex-paper-en skill for final formatting, consistency, and compilation check.

Step 8.4: Generate Final Package

Produce these deliverables in the project directory:

research-output/
  phase1-plan.md           — Research plan
  phase2-deep-research.md  — Web research notes
  phase2-academic-researcher.md — Academic research notes
  phase2-medical-imaging.md     — (if MEDICAL) Domain notes
  phase2-merged.md         — Cross-validated synthesis
  phase3-draft.md          — Initial draft
  phase4-citation-report.md — Citation issues log
  phase6-factcheck.md      — Verification report
  phase6-corrections.md    — Changes made
  phase7-peerreview.md     — Peer review report

manuscript.tex             — Final LaTeX manuscript
references.bib             — Validated BibTeX database
VERIFICATION_STATUS.md     — Overall confidence + caveats

Step 8.5: Final Summary

Report to user:

• Word count, source count, reference count
• Fact-check confidence level
• Peer review recommendation
• Any remaining unverified claims or caveats
• How to compile (pdflatex / xelatex / lualatex command)

GATE 5: Final delivery.

Ask: "All deliverables are ready in research-output/. Would you like me to: (a) walk through the verification report, (b) list any remaining unverified claims, or (c) deliver as final?" This is the last checkpoint — the user signs off on the complete package.

---

Anti-Patterns

Don't	Because	Do Instead
Ask agents to write files	Background agents often lack Write/Bash permissions	Agents return text inline; main session writes files
Combine Phase 3 + Phase 6	Same pass can't reliably catch own hallucinations	Always separate generation and verification
Skip Phase 2 merge	Duplicate findings ≠ wasted effort; they're cross-validation	Merge and note agreement/disagreement
Run only one research skill	Single source type → blind spots	At minimum deep-research + academic-researcher
Skip human gates	User should confirm direction and review drafts	Always gate at Phase 1, Phase 3, Phase 5
Let fact-check find errors → keep going	Errors compound	Fix all CONTRADICTED claims before Phase 7
Launch agents sequentially	They'd run one at a time, defeating the purpose of parallel	Always put all Phase 2 Agent calls in ONE message
Use Skill tool for Phase 2	Skill tool loads instructions into current context, not for delegation	Use Agent tool with tailored prompts for each perspective

Quality Heuristics

• Source diversity: Aim for ≥ 5 unique domains/hosts per research skill
• Recency: Prefer sources from last 3 years for fast-moving fields
• Citation completeness: Every factual claim traceable to a source
• Unverified ceiling: If > 15% claims are unverified, confidence must be Low
• Revision loop: If peer review says Major Revision, loop once; if still Major, flag to user

Quick Start (for the user)

# Session 1 — Research + Draft
# In any project directory, start Claude Code and type:
/academic-orchestrator

# Or say:
"Use the academic orchestrator to write a survey paper on [topic]"

# Follow Phase 1-3. When session ends, run the /compact command.
# All files are saved to research-output/.


# Session 2 — Citations + LaTeX
# In the same directory, start a new session:
/academic-orchestrator
# Then say: "continue from Phase 4"

# Or resume the previous session directly:
claude -r "previous-session-name" "continue the academic orchestrator from Phase 4"


# Session 3 — Verify + Review + Final
# In the same directory, start a new session:
/academic-orchestrator
# Then say: "continue from Phase 6"

# Or:
claude -c "continue the academic orchestrator from Phase 6"

The orchestrator reads intermediate files from research-output/ to pick up where it left off. Each session starts fresh (low context usage) and loads only what it needs from disk. Three sessions total, measured at ~50-75K, ~15-35K, and ~35-55K tokens respectively.

Markdown-only output: At Phase 1, say "Markdown only, skip LaTeX." Phase 5 is skipped but the 3-session split remains unchanged — Session 2 does citations only (shorter, ~15K), Session 3 does verification + review as normal. The session split is about context management, not LaTeX dependency.