Research write

Research Write Skill

Description

Orchestrates multi-agent collaborative writing from upstream research artifacts. Produces structured documents (papers, proposals) with evidence-grounded prose, MAGI cross-review, and automated quality validation.

Usage

/research-write --source <output_dir> [--mode paper|proposal] [--audience general-public|high-school|undergraduate|phd-student|researcher|expert|"free text"] [--depth low|medium|high] [--claude-only] [--resume <write_dir>]

Arguments

• $ARGUMENTS — Required and optional flags:
  • --source <output_dir> — Path to an upstream research output directory containing artifacts from /research, /research-brainstorm, /research-explain, or /research-report. Required unless --resume is provided.
  • --mode — Writing mode (default: paper):
    • paper — Academic research paper with standard structure (abstract, introduction, methodology, experiments, results, discussion, limitations, conclusion, references)
    • proposal — Grant or funding proposal with persuasive structure (executive summary, problem statement, proposed approach, preliminary results, timeline, budget, broader impact, references)
  • --audience — Target audience (default: researcher):
    • general-public — No assumed technical background
    • high-school — Basic math/science literacy
    • undergraduate — Introductory college-level knowledge in the domain
    • phd-student — Graduate-level domain knowledge
    • researcher — Active researcher familiar with the field (default)
    • expert — Deep specialist in the exact sub-field
    • "free text" — Any custom audience description (e.g., "medical doctors learning ML")
  • --depth — Controls review thoroughness (default: medium):
    • low — Skip MAGI cross-review; Claude reviews alone
    • medium — Standard MAGI cross-review (Gemini + Codex review in parallel)
    • high — MAGI cross-review + Devil's Advocate adversarial pass on high-stakes sections
  • --claude-only — Replace all Gemini/Codex MCP calls with Claude Agent subagents. Use when external model endpoints are unavailable. See Claude-Only Mode section below.
  • --resume <write_dir> — Resume an interrupted write pipeline from a previous write output directory. See Resume Protocol below.

Instructions

Shared rules: Read ${CLAUDE_PLUGIN_ROOT}/shared/rules.md before starting. §MCP, §Claude-Only, §LaTeX, §Visualization, §Substitute apply to this skill. Inline fallback (if shared rules unavailable): Gemini models: gemini-3.1-pro-preview → gemini-2.5-pro → Claude. Codex: gpt-5.4. All math in LaTeX only (no Unicode: σ₁→$\sigma_1$). Use @filepath for MCP file refs; subagents use Read tool.

MCP Tool Rules

See §MCP, §Visualization, §LaTeX in shared rules. Additionally:

• Codex: Use mcp__codex-cli__ask-codex for analysis/review.
• When to search: citation verification, related work references, fact-checking claims, confirming state-of-the-art results, checking terminology conventions

Claude-Only Mode

See §Claude-Only and §Substitute in shared rules.

LaTeX Formatting Rules

See §LaTeX in shared rules.

When this skill is invoked, follow these steps exactly:

---

Phase 0: Setup & Intake

Step 0a: Parse Arguments

1. Parse $ARGUMENTS:

   • Extract --source <output_dir> (required unless --resume). Validate the directory exists using Glob.
   • Extract --mode (default: paper). Validate against supported modes: paper, proposal.
   • Extract --audience (default: researcher). Store as-is (predefined keyword or free-text string).
   • Extract --depth (default: medium). Validate: low, medium, high.
   • Extract --claude-only (boolean, default: false).
   • Parse --resume <write_dir>: If provided, skip to the Resume Protocol below.

2. Determine the domain by reading upstream artifacts:

   • Check for brainstorm/weights.json in the source directory — the _meta.domain field contains the domain.
   • If not found, check brainstorm/personas.md for domain context.
   • If neither exists, infer domain from the source directory name or contents.

3. Create the write output directory inside the source directory:

   {source_dir}/write/
   

   If write/ already exists, announce that existing artifacts will be preserved and new artifacts will be added alongside them.

4. Announce to the user: source directory, mode, audience, depth, domain, and claude-only status.

Step 0b: Locate Upstream Artifacts

Inventory the source directory for available artifacts using Glob:

Artifact	Path	Required	Purpose
Brainstorm synthesis	brainstorm/synthesis.md	Recommended	Research directions, key findings
Research plan	plan/research_plan.md	Recommended	Methodology, objectives
Murder board	plan/murder_board.md	Optional	Stress-test results
Mitigations	plan/mitigations.md	Optional	Plan revisions
Source code	src/**/*	Optional	Implementation details
Test results	tests/**/*	Optional	Validation data
Plot manifest	plots/plot_manifest.json	Optional	Figures with metadata
Report	report.md	Optional	Pre-existing report draft
Explain outputs	explain/**/*.md	Optional	Concept explanations
Brainstorm personas	brainstorm/personas.md	Optional	Expert persona context
Weights	brainstorm/weights.json	Optional	Scoring weights used

For each artifact found, read the first 5 lines to confirm it is non-empty.

On missing recommended artifacts: Warn the user which recommended artifacts are absent. Ask: "The following recommended artifacts are missing: [list]. The document will have thinner coverage in the corresponding sections. Proceed anyway?" Continue on confirmation.

Step 0c: Generate Intake Artifacts

Claude reads all located upstream artifacts and generates two structured JSON files. This is an LLM extraction task — Claude performs the semantic extraction; Python validates the schema.

Schema reference: Read references/intake_schemas.md for the full write_inputs.json and citation_ledger.json schemas and extraction guidelines.

1. Generate write/write_inputs.json using the schema in references/intake_schemas.md.

2. Generate write/citation_ledger.json using the schema in references/intake_schemas.md.

3. Validate with maintained utility:

Determine the plugin root directory by navigating two levels up from this skill's base directory (e.g., if this skill is loaded from .../skills/research-write/, the plugin root is ../../ relative to that path). The Base directory for this skill: header injected by Claude Code provides the absolute path. Then run:

uv run python <plugin_root>/utils/validate_intake.py write/write_inputs.json write/citation_ledger.json

Important: Do NOT generate a validation script at runtime. Use the maintained utils/validate_intake.py utility.

If validation fails, Claude fixes the JSON files and re-runs validation. Maximum 2 fix attempts.

---

Phase 1: Outline

Step 1a: Load Mode Template

Read the mode template from ${CLAUDE_PLUGIN_ROOT}/templates/writing/{mode}.md.

Parse the YAML frontmatter to extract:

• sections — ordered list of section definitions with id, required, max_words, evidence_slots, style, narrative_role
• export — target export formats
• total_max_words — overall word budget
• tone, jargon_budget, formality — style constraints

Read the Markdown body for section dependencies and evidence integration guidelines.

Save a copy of the loaded template to write/mode_template_cache.md before proceeding to Step 1b. This cached copy is referenced by @filepath in subsequent model calls.

Step 1b: MAGI Parallel Outline Generation

Full prompts: Read references/outline_prompts.md for the complete MCP call specifications. Execute both calls simultaneously.

Generate two independent outlines using the mode template, write_inputs.json, and audience context:

• Gemini (BALTHASAR) — Creative Outline emphasizing narrative flow and reader engagement. Save to write/gemini_outline.md.
• Codex (CASPER) — Structural Outline emphasizing evidence coverage and completeness. Save to write/codex_outline.md.

If --claude-only: see the Claude-Only Fallback in references/outline_prompts.md.

Step 1c: Synthesize Section Contracts

Claude reads both outlines and synthesizes a canonical outline with per-section contracts:

1. Merge outlines: For each section, take the stronger content from each outline:

   • Prefer Gemini's narrative framing and transitions
   • Prefer Codex's evidence coverage and structural completeness
   • Resolve conflicts by favoring the version that covers more evidence items

2. Define the global argument thread: Synthesize both outlines' argument threads into a single 3-5 sentence narrative arc that connects all sections.

3. Generate write/section_contracts.json using the schema in references/intake_schemas.md.

4. Determine drafting order: Based on section dependencies from the mode template:

   • Independent sections first (e.g., related_work for papers, problem_statement for proposals)
   • Dependent sections next, respecting dependency chains
   • Summary sections last (e.g., abstract for papers, executive_summary for proposals)
   • The drafting_order field is a 1-indexed integer reflecting this order.

5. Evidence coverage check: Verify that every high-confidence claim from write_inputs.json appears in at least one section's claim_ids. Verify that every evidence item appears in at least one section's evidence_ids. Report any orphaned claims or evidence.

6. Save the synthesized outline to write/outline.md (human-readable Markdown) and write/section_contracts.json (machine-readable).

Step 1d: User Checkpoint — Outline Approval

>>> USER CHECKPOINT: Approve outline <<<

Present to the user:

• The global argument thread
• A summary table of all sections: title, purpose, narrative role, word budget, evidence count
• Any orphaned claims or evidence items
• The proposed drafting order

Wait for user approval. The user may:

• Approve the outline as-is
• Request modifications to specific sections (update section_contracts.json accordingly)
• Add or remove sections (within mode template constraints — required sections cannot be removed)
• Reorder the narrative

This is Hard Gate 1 — the most important checkpoint. No drafting begins until the user approves the outline.

---

Phase 2: Draft

Step 2a: Prepare Evidence Blocks

Before generating any section prose, Claude pre-assembles evidence blocks for each section. This is the core evidence integration strategy: LLMs write prose AROUND pre-placed evidence, NOT with macros.

For each section (in drafting order):

1. Read the section contract from write/section_contracts.json.
2. For each evidence_id in the section's evidence_ids, generate the appropriate block:
   • type == "plot":

     <!-- EVIDENCE BLOCK: ev-1 -->
     ![{caption}]({ref})
     *{caption}*
     <!-- END EVIDENCE BLOCK -->
     

     Anti-pattern: Do NOT list figures in a table at the end of the document. Every figure must be embedded inline with ![caption](path) immediately before or after the paragraph that discusses it.

   • type == "metric":

     <!-- EVIDENCE BLOCK: ev-2 -->
     **Key metric**: {description} — {value}
     <!-- END EVIDENCE BLOCK -->
     

   • type == "test_result":

     <!-- EVIDENCE BLOCK: ev-3 -->
     **Validation**: {description}
     <!-- END EVIDENCE BLOCK -->
     

   • type == "code":

     <!-- EVIDENCE BLOCK: ev-4 -->
     See implementation in `{ref}`: {description}
     <!-- END EVIDENCE BLOCK -->
     

3. Before embedding each plot evidence block, verify that the ref path exists (Glob check). If missing, mark as <!-- EVIDENCE BLOCK: {id} (MISSING FILE: {ref}) --> and log a warning.
4. Save pre-assembled evidence blocks to write/evidence_blocks/{section_id}.md.

Step 2b: Section-by-Section Drafting

Generate each section in the order specified by drafting_order in section_contracts.json. For each section:

Context provided (scoped to prevent context bloat):

• The section's own contract (from section_contracts.json)
• The global argument thread (2-3 sentences)
• Pre-assembled evidence blocks for this section (from write/evidence_blocks/{section_id}.md)
• The relevant claims from write_inputs.json (only claims listed in the section's claim_ids)
• Preceding section summaries: for each already-drafted section, provide only the section title + first paragraph + last paragraph
• Audience and style constraints from the mode template

For each section, Claude:

1. Reads the section contract, evidence blocks, and relevant claims
2. Writes the section prose, integrating pre-placed evidence blocks naturally into the narrative
3. Ensures the section:
   • Opens with a transition from the previous section (transition_from_previous field)
   • Covers all key points listed in the contract
   • References all pre-placed evidence with concrete quantitative observations (not passive "see figure below")
   • Stays within the max_words budget (±10%)
   • Matches the style, tone, and formality constraints
   • Closes with a transition to the next section (transition_to_next field)
4. After writing the section, generates a 1-2 sentence summary for use as context by subsequent sections

Save each section to write/sections/{section_id}.md. Save each section summary to write/sections/{section_id}_summary.md.

Step 2c: Assemble Full Draft

After all sections are drafted:

1. Concatenate all sections in the order defined by the mode template (NOT drafting order — presentation order).
2. Add a document title derived from the global argument thread and source directory topic.
3. For paper mode: add author placeholder, date, and abstract positioning.
4. For proposal mode: add a title page with project title, PI placeholder, and date.
5. Save the assembled draft to write/draft.md.

Step 2d: Narrative Arc Tracking

Claude reads the full assembled draft and evaluates the narrative arc:

1. Thread continuity check: Does each section's opening connect to the previous section's closing? Are there abrupt topic shifts?
2. Argument progression: Does the document build its case progressively? Does the conclusion follow from the evidence presented?
3. Tone consistency: Is the voice consistent across all sections? (Flagged only — resolution happens in Phase 3.)

Save the narrative arc assessment to write/narrative_arc_assessment.md. This feeds into the Phase 3d global coherence pass.

---

Phase 3: Review

Step 3a: MAGI Cross-Review

Skip this step if --depth low. For --depth low, proceed directly to Step 3c (DocCI Validation).

Full prompts: Read references/review_prompts.md for the complete MCP call specifications. Execute both calls simultaneously.

Execute two parallel review calls:

• Gemini (BALTHASAR) — Content Quality Review: claim support, evidence integration, narrative flow, audience fit, completeness. Save to write/gemini_review.md.
• Codex (CASPER) — Structure & Evidence Review: word budget compliance, evidence completeness, citation integrity, LaTeX correctness, structural compliance. Save to write/codex_review.md.

If --claude-only: see the Claude-Only Fallback in references/review_prompts.md.

Step 3b: Devil's Advocate Review

Skip this step if --depth low or --depth medium. Only execute for --depth high.

Full spec: Read references/depth_high.md for selection criteria, the full adversarial prompt, Claude-Only fallback, and fix application rules.

Identify high-stakes sections (most critical/major issues from Step 3a + required sections with evidence/method/result narrative roles). Apply adversarial review to at most 3 sections. Save each result to write/devils_advocate_{section_id}.md.

Step 3c: Claude Synthesizes Reviews & Applies Fixes

Claude reads all review outputs and applies fixes:

1. Read reviews: write/gemini_review.md, write/codex_review.md, and any write/devils_advocate_*.md files.

2. Triage issues:

   • Consensus issues (flagged by both Gemini and Codex): High-priority fixes, apply immediately
   • Divergent issues (flagged by only one reviewer): Evaluate on merit, apply where appropriate
   • Devil's Advocate findings: Apply all Critical fixes; evaluate Major fixes on merit

3. Apply section-level fixes: For each section with issues:

   • Re-read the section contract and evidence blocks
   • Rewrite the section incorporating all accepted fixes
   • Ensure fixes don't introduce new problems (check word budget, evidence coverage)
   • Save the revised section to write/sections/{section_id}.md (overwrite)

   Data integrity: Do NOT fabricate data for illustrative plots. If no data exists for a needed visualization, note the gap textually rather than generating synthetic values.

4. Escalation trigger: If any fix requires changing a section's scope or adding/removing key points not in the section contract, update section_contracts.json and flag this for the user:

   "Section contract for '{section_id}' has been modified during review. Changes: [list]. These will be shown for approval in Phase 4."

5. Re-assemble the revised draft: write/revised_draft.md.

Step 3d: Global Coherence Pass

After all section-level fixes are applied, Claude performs a focused coherence pass:

1. Read the full write/revised_draft.md.
2. Read write/narrative_arc_assessment.md from Phase 2d.
3. Rewrite ONLY transition sentences — the first and last paragraphs of each section — to create narrative flow across the entire document.
4. Check that the global argument thread is maintained from introduction through conclusion.
5. Save the coherence-edited draft to write/revised_draft.md (overwrite).

This is NOT a full rewrite. It is a bounded, focused pass that touches only inter-section boundaries. The goal is to eliminate the "Frankenstein" effect of section-by-section generation.

Step 3e: DocCI Validation

Run the repository-maintained draft validator to perform automated quality checks.

Determine the plugin root directory by navigating two levels up from this skill's base directory. The Base directory for this skill: header injected by Claude Code provides the absolute path. Then run:

uv run python <plugin_root>/utils/validate_draft.py write/revised_draft.md write/section_contracts.json write/write_inputs.json

Important: Do NOT generate a validation script at runtime. Use the maintained utils/validate_draft.py utility.

On validation failure: Claude reads the validation report, fixes the identified issues in write/revised_draft.md, and re-runs validation. Maximum 2 fix iterations.

Save the validation report to write/validation_report.json.

---

Phase 4: Finalize

Step 4a: Macro Resolution Fallback

Before presenting to the user, scan the draft for any remaining unresolved macros. If MAGI model outputs introduced any {{fig:id}} or {{ref:id}} patterns during review, resolve them now:

1. Scan write/revised_draft.md for patterns matching {{fig:...}} or {{ref:...}}.
2. If any are found, read references/macro_resolution.md for the full script source, write it to write/resolve_macros.py, then run:

   uv run python write/resolve_macros.py write/revised_draft.md write/write_inputs.json write/citation_ledger.json
   

   If unresolved macros remain, Claude manually resolves them by reading the intake data and making inline replacements.
3. If no macros are found, skip this step.

Step 4b: User Checkpoint — Final Approval

>>> USER CHECKPOINT: Approve final document <<<

Present to the user:

• The final draft location: write/revised_draft.md
• A summary table:
  • Total word count and per-section word counts vs. budgets
  • Number of evidence items integrated
  • Number of claims supported vs. unresolved
  • DocCI validation status (pass/fail/warnings)
  • Review issues found and resolved (counts by severity)
• Any section contract modifications from Phase 3 (if escalation was triggered)
• Any remaining warnings from DocCI validation

The user may:

• Approve the document as final
• Request specific section revisions (Claude revises and re-validates)
• Request a full re-review at a different --depth level

This is Hard Gate 2 — the final publish gate. No export happens until the user approves.

Step 4c: Export

After user approval:

1. Copy the approved draft to the final document name:

   • Paper mode: write/{topic}_paper.md
   • Proposal mode: write/{topic}_proposal.md

2. Clean up evidence block markers: Remove all <!-- EVIDENCE BLOCK: ... --> and <!-- END EVIDENCE BLOCK --> HTML comments from the final document, leaving only the rendered content.

3. Generate write/writing_state.json using the schema in references/intake_schemas.md.

4. Present completion summary to the user:

   • Final document path
   • Word count and section breakdown
   • Evidence integration statistics
   • Any remaining caveats or limitations

---

Resume Protocol

When --resume <write_dir> is provided, the pipeline skips initialization and infers the current phase from the presence of key artifact files in the write output directory. The artifacts themselves serve as checkpoints — no separate state file is required for resume inference.

Phase inference rules (evaluated top-down; first match wins):

Condition	Inference	Action
write/writing_state.json exists with status: "complete"	Pipeline complete	Inform user; offer to re-run specific phases
write/revised_draft.md exists	Phase 3 complete	Resume from Phase 4 (Finalize)
write/draft.md exists	Phase 2 complete	Resume from Phase 3 (Review)
write/section_contracts.json exists	Phase 1 complete	Resume from Phase 2 (Draft)
write/write_inputs.json exists	Phase 0 complete	Resume from Phase 1 (Outline)
None of the above	No phase complete	Start from Phase 0 (Setup & Intake)

Resume procedure:

1. Use Glob to check for each artifact in the order above.
2. Read the first few lines of the matched artifact to confirm it is non-empty.
3. Announce to the user: detected phase, write directory, and which phase will be resumed.
4. Infer --source, --mode, and --audience from write/write_inputs.json if it exists. If not, require the user to provide --source.
5. Continue the pipeline from the inferred phase, skipping all prior phases.

Important: On resume, do NOT re-create the write directory or overwrite existing artifacts from prior phases. Only create artifacts for the resumed phase and beyond.

---

Output Files

The write pipeline produces artifacts in the {source_dir}/write/ directory:

{source_dir}/
└── write/
    ├── write_inputs.json            # Canonical intake (claims, evidence, definitions)
    ├── citation_ledger.json         # Claim-source tracking
    ├── mode_template_cache.md       # Cached copy of mode template
    ├── gemini_outline.md            # Gemini's outline proposal
    ├── codex_outline.md             # Codex's outline proposal
    ├── outline.md                   # Synthesized canonical outline (human-readable)
    ├── section_contracts.json       # Per-section contracts (machine-readable)
    ├── evidence_blocks/
    │   ├── introduction.md          # Pre-assembled evidence for each section
    │   ├── methodology.md
    │   ├── results.md
    │   └── ...
    ├── sections/
    │   ├── introduction.md          # Individual section drafts
    │   ├── introduction_summary.md  # Section summary for context scoping
    │   ├── methodology.md
    │   ├── methodology_summary.md
    │   └── ...
    ├── draft.md                     # Assembled first draft
    ├── narrative_arc_assessment.md  # Narrative arc evaluation
    ├── gemini_review.md             # Gemini content quality review
    ├── codex_review.md              # Codex structure & evidence review
    ├── devils_advocate_*.md         # Devil's Advocate reviews (--depth high only)
    ├── revised_draft.md             # Draft after review fixes + coherence pass
    ├── validation_report.json       # DocCI validation results
    ├── resolve_macros.py            # Macro resolution fallback script (if needed)
    ├── {topic}_paper.md             # Final export (paper mode)
    ├── {topic}_proposal.md          # Final export (proposal mode)
    └── writing_state.json           # Pipeline state manifest

Notes

• The write skill is designed as a standalone skill invoked after upstream research phases are complete. It is NOT embedded within the /research pipeline — invoke it separately with --source pointing to a research output directory.
• If upstream artifacts are incomplete, the skill will produce a document with thinner coverage in the corresponding sections rather than failing entirely. Missing sections are flagged in the DocCI validation.
• The evidence-first approach (pre-inserting evidence blocks before prose generation) prevents hallucination by ensuring the LLM writes around grounded data rather than fabricating evidence to fit a narrative.
• Validation utilities (validate_intake.py, validate_draft.py) are maintained in the utils/ directory at the plugin root. They are NOT generated at runtime — resolve the plugin root by navigating two levels up from this skill's base directory and use the maintained versions at <plugin_root>/utils/. The resolve_macros.py script is still written by Claude during the pipeline when needed.
• The two hard gates (outline approval + final publish) ensure the user retains authorial control over structure and content. Between gates, the pipeline progresses automatically with review-driven quality assurance.
• Section-by-section drafting uses scoped context windows: each section receives only its contract, evidence, and preceding section summaries — not the full document or full intake data. This prevents context bloat and "lost in the middle" failures.
• The global coherence pass (Phase 3d) rewrites ONLY transition sentences (first/last paragraphs of each section). It is NOT a full rewrite — it is a bounded task that eliminates the "Frankenstein" effect.
• For mathematical content, include LaTeX formatting instructions in MAGI model prompts. All mathematical expressions must follow the LaTeX Formatting Rules section above.
• When writing about plots and figures, always include concrete quantitative observations — never just "As shown in Figure X."
• The mode template at ${CLAUDE_PLUGIN_ROOT}/templates/writing/{mode}.md defines section structure, word budgets, evidence slots, and style constraints. Adding new modes requires only a new template file.