draft-review

Draft Review Skill

A multi-agent academic paper review system that provides comprehensive feedback across multiple dimensions.

Usage

/draft-review path/to/paper.pdf
/draft-review path/to/paper.tex --code path/to/code/

Review Workflow

Phase 1: Document Ingestion

1. PDF Input: Convert PDF to markdown using mistral-pdf-to-markdown skill
2. TeX Input: Read tex file directly for mathematical precision
3. Code (Optional): Index relevant code files if path provided

After ingestion, create a document structure summary:

• Paper title and abstract
• Section structure with approximate lengths
• List of figures and tables with captions
• Key notation/variable definitions index

Phase 2: Review Configuration

Use AskUserQuestion to configure the review:

Question 1: Review Scope

header: "Scope"
question: "Which aspects should I review?"
options:
  - label: "Comprehensive (Recommended)"
    description: "All review aspects: math, writing, consistency, arguments, proofreading, citations"
  - label: "Mathematical"
    description: "Focus on derivations, equations, proofs, and notation"
  - label: "Writing & Clarity"
    description: "Writing quality, terminology consistency, and structure"
  - label: "Quick Proof"
    description: "Typos, grammar, and formatting only"
multiSelect: false

Question 2: Thoroughness Level

header: "Thoroughness"
question: "How thorough should the review be?"
options:
  - label: "Standard (Recommended)"
    description: "Thorough single-agent review per category"
  - label: "Quick"
    description: "Fast surface-level review"
  - label: "Deep"
    description: "Multiple parallel agents per category with diverse perspectives for maximum coverage"
multiSelect: false

Phase 3: Dispatch Subagents

Based on configuration, dispatch appropriate subagents using the Task tool:

Scope	Agents to Dispatch
Comprehensive	All 6 (or 7 if code provided)
Mathematical	mathematical-reviewer only
Writing & Clarity	writing-clarity-reviewer only
Quick Proof	proofreader only

For each subagent, provide:

1. Document Summary (~500 tokens): Title, abstract, section structure
2. Relevant Sections (~2000-5000 tokens): Only sections pertinent to that agent's focus
3. Cross-Reference Index: Tables, figures, key definitions

Example Context for Mathematical Reviewer

Document: [title], [abstract summary]
Sections to Review: Appendix A, Appendix B, Section 3 (Model)
Notation Index: {ζ: elasticity, σ²: variance, β: coefficient, ...}
Tables Referenced: Table 1 (main estimates), Table B.3 (robustness)

Phase 4: Deep Mode - Parallel Agent Strategy

When thoroughness is "Deep", run 2-3 agents per category with diverse perspectives:

Perspective Variations:

• Agent A: "Review as a skeptical referee looking for flaws"
• Agent B: "Review as a constructive mentor suggesting improvements"
• Agent C: "Review as a domain expert in [specific methodology]"

Focus Ordering Variations:

• Agent A: Start from beginning, work forward
• Agent B: Start from conclusions, trace claims backward
• Agent C: Start from most complex/critical section

After parallel runs, merge findings:

• Deduplicate similar issues
• Flag issues found by multiple agents as higher confidence
• Include unique findings (may catch edge cases)

Phase 5: Result Aggregation

Collect all subagent outputs and organize by severity:

Critical (Priority 1):

• Mathematical errors in proofs/derivations
• Contradictory claims
• Missing critical references
• Data inconsistencies affecting results

Major (Priority 2):

• Logical gaps in argumentation
• Unclear methodology descriptions
• Significant notation inconsistencies
• Writing clarity issues affecting comprehension

Minor (Priority 3) - Potentially Auto-fixable:

• Typos and grammatical errors
• Minor formatting issues
• Small notation inconsistencies
• Reference format issues

Phase 6: Task Generation (User Permission Required)

Present summary to user:

Review complete. Found:
- X Critical issues
- Y Major issues
- Z Minor issues (N auto-fixable)

Use AskUserQuestion:

header: "Tasks"
question: "Would you like me to create tasks for auto-fixable minor issues?"
options:
  - label: "Yes, create tasks (Recommended)"
    description: "I'll create TodoWrite tasks for typos, formatting, and other quick fixes"
  - label: "No, show report only"
    description: "Just show the full review report without creating tasks"
multiSelect: false

If approved, use TodoWrite to create tasks for each auto-fixable issue.

Subagent Dispatch Instructions

Use the Task tool with the appropriate registered subagent_type for each review agent.

Available Subagent Types

Subagent Type	Purpose
draft-reviewer:mathematical-reviewer	Verify derivations, proofs, equations, notation
draft-reviewer:writing-clarity-reviewer	Writing quality and clarity
draft-reviewer:consistency-checker	Internal consistency of claims, numbers, terminology
draft-reviewer:argument-logic-reviewer	Logical flow and argumentation
draft-reviewer:proofreader	Typos, grammar, formatting
draft-reviewer:citation-checker	Citation completeness and accuracy
draft-reviewer:code-paper-consistency	Verify code matches paper claims (if code provided)

Dispatch Template

Task tool parameters:
  subagent_type: "draft-reviewer:[agent-name]"
  description: "[Agent type] review"
  prompt: |
    Review the following academic paper sections.

    Document Summary:
    [Insert summary]

    Sections to Review:
    [Insert relevant sections]

    Cross-Reference Index:
    [Insert index]

    Output your findings using this format:
    ### [SEVERITY] [Category]: [Brief Title]
    **Location:** [Section/equation/page]
    **Issue:** [Description]
    **Recommendation:** [Suggested fix]
    **Auto-fixable:** [Yes/No]

Scope to Agent Mapping

Subagent Type	Comprehensive	Mathematical	Writing	Quick
draft-reviewer:mathematical-reviewer	✓	✓
draft-reviewer:writing-clarity-reviewer	✓		✓
draft-reviewer:consistency-checker	✓
draft-reviewer:argument-logic-reviewer	✓
draft-reviewer:proofreader	✓			✓
draft-reviewer:citation-checker	✓
draft-reviewer:code-paper-consistency	✓*

*Only if code path provided

Output Format

Final report structure:

# Draft Review Report: [Paper Title]

## Summary
- Total issues found: X
- Critical: X | Major: X | Minor: X
- Review scope: [scope]
- Thoroughness: [level]

## Critical Issues
[List with full details and recommendations]

## Major Issues
[List with details and suggestions]

## Minor Issues
[List with specific corrections]

## Auto-Fixable Items
[List of items that can be addressed via tasks]

Dependencies

• mistral-pdf-to-markdown skill (for PDF input)
• Task tool with general-purpose subagent
• TodoWrite tool (for task generation)
• AskUserQuestion tool (for configuration)
• Read, Glob, Grep tools (for document processing)

Notes

• Each review session focuses on one paper for thorough analysis
• For papers >40 pages, the hierarchical context management is critical
• Always read the full tex source if available (more precise than PDF conversion)
• Code-paper consistency check requires explicit code path