Skill: paper-structure-planning

A systematic approach to planning and revising academic paper structure through detailed outline and paragraph-level analysis.

Overview

This skill enables structured paper planning where:

•Claude creates hierarchical structure (outline → paragraph-level)
•User reviews and adds modification comments
•Claude applies inline edits with strikethrough formatting
•All planning tracked in a logging file (NOT direct LaTeX edits)

The logging file becomes the blueprint for paper writing/revision.

When to Use

•Planning structure for a new paper
•Revising rejected paper for different venue
•Analyzing existing paper paragraph-by-paragraph
•Learning from sample papers to create new structure
•Strategic repositioning of content across papers

Trigger phrases:

•"plan the paper structure for..."
•"analyze the paper structure of..."
•"revise this paper structure for [venue]..."
•"create paper outline based on these sample papers..."

File Structure

Two-Level Hierarchy:

code

Session 1: Paper Outline (YYYY-MM-DD)
======================================

Overview
--------

{High-level description of paper goals, venue fit, scope}

**Target Audience**: {who will read this}

**Venue Fit**: {why this venue, what they value}

**Scope & Structure (X pages)**:


Abstract
--------

- Problem: {1-2 lines}
- Solution: {1-2 lines}
- Contribution: {1-2 lines}


Introduction (X pages)
----------------------

- {Bullet point outline}
- {Key narrative arc}
- {Main contributions}


Section 2: {Title} (X pages)
-----------------------------

- {Content outline}
- {Key points to cover}


Discussion/Limitations/Conclusion
----------------------------------

- {What to emphasize}
- {What to omit}


**Key Omissions** (moved to Paper B or out of scope):
- NO {topic that doesn't fit}
- NO {analysis that's too deep}



Session 2: Paragraph-Level Structure (YYYY-MM-DD)
==================================================

Overview
--------

Detailed paragraph-by-paragraph breakdown of paper structure.
Every paragraph catalogued with purpose and key claims.


Section Title (Line X-Y)
-------------------------

**P1 (line X)**: {One-sentence summary of paragraph content}

**P2 (line Y)**: {One-sentence summary}

**Subsection (line A-B)**:

**P3**: {Summary}


{Continue for entire paper...}



Session 3: Appendix/Supplementary Structure
============================================

{Same paragraph-level format for appendix}

Key Principles

1. Paragraph-Level Granularity

Every paragraph gets:

•Sequential number (P1, P2, P3...)
•Line number reference (if analyzing existing paper)
•One-sentence summary of content/purpose
•Key claims or data points

2. Outline-First Approach

Session 1 establishes:

•Page budgets per section
•Scope boundaries (what's IN vs OUT)
•Strategic positioning for venue
•Content distribution decisions

3. Inline Modification Format

When user requests changes, use strikethrough with two-line format:

code

**P1**: ~~Old content description~~
        New content description >> CC-modify

**P2**: ~~Text to delete~~ >> CC-modify: DELETE

**P3 (NEW)**: New paragraph to add >> CC-modify

Format rules:

•KEEP = no comment (leave original)
•MODIFY = ~~old~~ on first line, new text on second line (indented) >> CC-modify
•DELETE (single line) = ~~text~~ >> CC-modify: DELETE
•DELETE (multi-line) = Each line gets ~~ markers, then >> CC-modify at end
•ADD = new text >> CC-modify (mark as NEW)

Readability tip: For long modifications, always use two lines (strikethrough, then new text indented with spaces)

Multi-line deletion example:

code

~~**Section Title**:~~
~~**P1**: First point to delete~~
~~**P2**: Second point to delete~~
>> CC-modify: DELETE

4. Strategic Scope Management

Clearly document what's excluded:

•"Key Omissions" section in Session 1
•Explicit DELETE markers for out-of-scope content
•Redirect to other papers (Paper B, future work)

5. Venue-Specific Positioning

Tailor structure to venue expectations:

•Conference track (Datasets & Benchmarks vs Main Track)
•Page limits and appendix rules
•What reviewers value (infrastructure vs insights)
•Positioning statement (benchmark paper vs analysis paper)

Workflow

Case 1: New Paper from Scratch

code

User: "Plan structure for ICML paper on FairGlucose benchmark"

Step 1: Create Session 1 - Outline
→ Claude writes outline with scope, page budgets
→ User reviews, adds "> JL:" comments

Step 2: Create Session 2 - Paragraph Structure
→ Claude expands to paragraph-level (P1, P2, P3...)
→ User reviews, requests modifications

Step 3: Apply Modifications
→ Claude uses ~~strikethrough~~ format
→ User approves → ready to write paper

Case 2: Revise Rejected Paper for New Venue

code

User: "This AAAI paper was rejected. Revise for ICML Datasets track."

Step 1: Analyze Existing Structure (Session 2)
→ Claude catalogs every paragraph with line numbers
→ Identifies what worked, what didn't
→ Maps rejection feedback to structure

Step 2: Create New Outline (Session 1)
→ Strategic repositioning for new venue
→ Content transformation plan (DELETE X, ADD Y)
→ Scope adjustment (what moves to Paper B)

Step 3: Modify Paragraph Structure
→ Inline edits with strikethrough
→ Track additions/deletions/modifications
→ User approves → ready to revise paper

Case 3: Learn from Sample Papers

code

User: "Analyze these 3 benchmark papers and plan structure for mine"

Step 1: Analyze Samples (Session 2, 3, 4...)
→ Claude catalogs paragraph structure of each sample
→ Identifies common patterns
→ Notes effective vs ineffective elements

Step 2: Synthesize into New Outline (Session 1)
→ Best practices from samples
→ Adapted to user's specific contribution
→ Venue-appropriate structure

Step 3: Expand to Paragraph Level
→ New Session with paragraph breakdown
→ User reviews and refines

Session Guidelines

Session 1: Outline Level

Purpose: High-level structure and scope decisions

Required sections:

•Overview (audience, venue fit, scope)
•Abstract structure
•Section-by-section outline with page budgets
•Key Omissions (what's excluded and why)

Keep it concise: 1-2 pages max

Session 2+: Paragraph Level

Purpose: Detailed content planning

Format for each paragraph:

code

**P{N} ({line X} if applicable)**: {One-sentence summary}
  - {Optional: key data points}
  - {Optional: specific claims to make}

Group by sections and subsections from Session 1

Session N: Modification Tracking

When revising, add summary section:

code

Summary of Changes
------------------

>> CC-modify: **CONTENT TRANSFORMATION**:
>> CC-modify: DELETE: ~X paragraphs (topics: A, B, C)
>> CC-modify: ADD: ~Y paragraphs (topics: D, E, F)
>> CC-modify: MODIFY: ~Z paragraphs (reframing/strengthening)

>> CC-modify: **NET EFFECT**:
>> CC-modify: - Section X: 8 para → 6 para (streamlined)
>> CC-modify: - Section Y: 5 para → 9 para (expanded events)
>> CC-modify: - Overall: 45 para → 42 para

Comment Conventions

User Comments:

code

> JL: {instruction or question}
> JL2: {follow-up comment}
> JL 3: {another point on same paragraph}

Claude Responses:

MUST respond to every user comment with:

code

>> CC: {action taken or answer}
>> CC-modify: {inline modification explanation}
>> CC2: {follow-up response}

Status Markers:

•DONE - Change completed
•DELETE - Content removed
•ADD - New content added
•MODIFY - Content revised
•KEEP - No change needed

Inline Editing Format

Original paragraph:

code

**P1**: Problem - CGM benchmarks have unbalanced datasets

User requests change:

code

> JL: Add event data gap here

Claude applies modification:

code

**P1**: ~~Problem - CGM benchmarks have unbalanced datasets~~
        Problem - CGM benchmarks have unbalanced datasets and lack event annotations >> CC-modify

For single-line deletions:

code

**P3**: ~~Input length sensitivity analysis~~ >> CC-modify: DELETE

For multi-line deletions (each line needs ~~):

code

~~**Subsection Title**:~~
~~**P1**: First paragraph to delete~~
~~**P2**: Second paragraph to delete~~
~~**P3**: Third paragraph to delete~~
>> CC-modify: DELETE

For additions:

code

**P5 (NEW)**: Event characterization - meal/exercise/medication frequencies across subgroups >> CC-modify

Strategic Positioning Patterns

Pattern 1: Infrastructure vs Insights

Infrastructure Paper (Benchmark, Dataset):

•Emphasize: reproducibility, accessibility, baseline coverage
•De-emphasize: deep analysis, novel insights, interpretations
•Structure: Dataset Design > Protocol > Baselines > Brief Results

Analysis Paper (Main Track):

•Emphasize: findings, insights, explanations
•De-emphasize: implementation details, baseline comparisons
•Structure: Motivation > Analysis > Insights > Implications

Pattern 2: Scope Partitioning

When content doesn't fit one paper:

code

Paper A (ICML Benchmark):
- Dataset + Protocol + Baselines
- OMIT: input length analysis, difficulty analysis

Paper B (Separate Analysis):
- Deep fairness analysis
- OMIT: dataset details (reference Paper A)

Pattern 3: Content Transformation

Common transformations when revising:

•DELETE interpretive claims → ADD factual reporting
•DELETE deep analysis sections → ADD infrastructure details
•MODIFY ambitious findings → baseline establishment
•ADD missing contributions (e.g., events in our case)

Output Format

Always produce a LOGGING FILE, never direct LaTeX edits.

File naming: vYYMMDD-{descriptor}-paper-structure.md

Example: v0116-icml-paper-structure.md

This file is:

•Planning document (blueprint for writing)
•Discussion space (with > JL and >> CC comments)
•Modification tracker (strikethrough edits)
•Historical record (why decisions were made)

NOT for:

•Direct paper writing
•LaTeX code generation
•Figure/table creation

Quality Checklist

Before finalizing structure, verify:

Outline Level (Session 1):

• Clear venue fit statement
• Page budgets for all sections
• Explicit scope boundaries (Key Omissions)
• Strategic positioning rationale
• Abstract structure with problem/solution/contribution

Paragraph Level (Session 2+):

• Every paragraph numbered and summarized
• Line numbers if analyzing existing paper
• Grouped by sections matching outline
• Key claims/data identified
• Logical flow within and between sections

Modifications:

• All user comments have >> CC responses
• Inline edits use ~~strikethrough~~ format
• Transformation summary provided
• Net paragraph count changes tracked

Strategic Clarity:

• Clear distinction what's IN vs OUT
• Content redirected if doesn't fit (Paper B, appendix)
• Positioning appropriate for venue
• No scope creep or ambition mismatch

Examples

Example 1: New Paper Planning

code

Session 1: ICML 2026 Benchmark Paper Outline (2026-01-16)
==========================================================

**Target Audience**: ML researchers, benchmark developers

**Venue Fit**: ICML Datasets & Benchmarks Track values infrastructure
contributions over algorithmic novelty

**Scope (8 pages)**: Dataset + Protocol + Baselines (NO deep analysis)

Abstract
--------
- Problem: Existing benchmarks lack balanced subgroups + event data
- Solution: FairGlucose with 300 patients, 12 strata, event annotations
- Contribution: Infrastructure enabling fairness research

Introduction (1.5 pages)
-------------------------
- CGM forecasting landscape
- Gap: no fairness-aware benchmark with events
- FairGlucose design principles
- Contributions: (1) dataset, (2) protocol, (3) baselines, (4) code

**Key Omissions** (moved to Paper B):
- NO input length sensitivity analysis
- NO instance difficulty analysis
- NO deep fairness insights

Example 2: Modification with Strikethrough (Two-Line Format)

code

Session 2: Paragraph-Level Structure
=====================================

Introduction
------------

**P7**: ~~Key findings - (a) PatchTST best; (b) LLMs robust on hard cases;
        (c) single-axis fairness misleading; (d) input length effects vary~~
        Baseline evaluation shows: (a) neural models achieve best accuracy;
        (b) subgroup disparities exist; (c) events correlate with challenges
        >> CC-modify

> JL: Good, remove the deep insights, keep factual baseline summary
>> CC: DONE. Removed interpretive findings (LLM properties, input length),
       replaced with basic baseline observations appropriate for benchmark paper.

Note: Strikethrough on first line, new text on second line (indented) with >> CC-modify marker. This format makes changes easy to review.

Anti-Patterns to Avoid

❌ Don't:

•Skip Session 1: Never jump straight to paragraph-level without outline
•Mix planning and writing: Logging file is for STRUCTURE, not prose
•Leave comments unresponded: Every > JL needs >> CC response
•Forget line numbers: When analyzing existing paper, always reference lines
•Ignore scope creep: If adding content, check against Session 1 scope
•Over-detail methods: Structure file summarizes paragraphs, doesn't write them
•Miss transformation summary: When revising, always track ADD/DELETE/MODIFY counts

✅ Do:

•Start with outline: Session 1 sets strategic direction
•Be granular: Every paragraph gets a number and summary
•Use strikethrough: Inline edits make changes crystal clear
•Track transformations: Count paragraphs added/deleted/modified
•Respect scope: Explicitly document what's out of scope
•Respond to all comments: Complete the conversation loop
•Think strategically: Position for venue, don't just list content

Integration with Other Skills

Works well with:

•coding-by-logging: Same commenting style (> JL, >> CC)
•Standard paper writing workflow: This skill produces blueprint → then write LaTeX

Workflow sequence:

•This skill → Create structure logging file
•User approves structure
•Write actual paper (LaTeX) following the structure
•Use coding-by-logging for revisions if needed

Tips for Effective Use

For New Papers:

•Start broad (Session 1), then drill down (Session 2)
•Reference sample papers from same venue/track
•Budget pages carefully (realistic about scope)
•Define "Key Omissions" upfront to avoid scope creep

For Revisions:

•Analyze rejection feedback first
•Map feedback to specific paragraphs
•Identify what to salvage vs delete
•Create transformation plan before editing

For Strategic Positioning:

•State positioning explicitly ("benchmark paper" vs "analysis paper")
•Check venue guidelines for what's valued
•Don't try to be two types of papers at once
•Redirect out-of-scope content to Paper B or future work

For Collaboration:

•User adds > JL comments after reviewing each session
•Claude responds with >> CC for every comment
•Iterate until structure is approved
•Structure file becomes shared blueprint for writing

End of Skill Definition