paper-writing

name	paper-writing
description	Assemble the full research paper from all upstream outputs into a single md file. Use when user says "写论文", "write paper", "assemble paper", or after all upstream outputs are ready.
argument-hint	["paper title or reference to idea01.md"]
allowed-tools	Bash(*), Read, Write, Edit, Grep, Glob, WebSearch, WebFetch, Agent

Skill-6: Paper Writing

Assemble the full research paper for: $ARGUMENTS

Overview

This skill integrates ALL upstream outputs into a single markdown file — the complete research paper. It reads existing outputs (idea, review, methodology, results, evaluation), embeds relevant tables directly into the text, writes new content for Discussion and Conclusion, and generates the Abstract last.

What is integrated (existing content, adapted into paper sections):

• Introduction framework → from idea01.md
• Literature Review → from literature_review.md (skill-5)
• Methodology → from methodology.md (SubAgent2)
• Results → from resultreport.md (SubAgent2) + evaluation.md (SubAgent3)

What is newly written by this skill:

• Discussion (based on results + evaluation + idea, with insights per Q1 standards)
• Conclusion (contributions, limitations, future outlook)
• Abstract (written last, ≤400 words)

Figures: Placeholders only ([Figure N: description]). Actual figures generated by skill-8.

Inputs

READ all of these before starting:

File	Section It Feeds
output/ideas/idea01.md	Introduction framework, research question, innovation points
output/papers/literature_review.md	Literature Review section
output/papers/review_tables/*.csv	Tables embedded in review section
output/ontology/ontology03.md	Methodology — ontology tables
output/subagent2_kg_graphrag/methodology.md	Methodology section
output/subagent2_kg_graphrag/resultreport.md	Results section
output/subagent3_evaluation/evaluation.md	Results section (evaluation part)
output/subagent3_evaluation/eval1_extraction/*.csv	Results — extraction tables
output/subagent3_evaluation/eval2_kg_quality/*.csv	Results — KG quality tables
output/subagent3_evaluation/eval3_rag_retrieval/*.csv	Results — retrieval comparison table
output/subagent3_evaluation/eval4_qa_quality/*.csv	Results — QA metrics table
output/subagent3_evaluation/eval5_llm_comparison/*.csv	Results — LLM comparison table
output/plans/plan04.md	Phase H task assignments

Output

Single file: output/papers/paper_draft.md

This file contains the entire paper in one markdown document, including all embedded tables.

Additionally: figure prompts to output/figure_prompts/paper/

References

Reference	Path	Purpose
Discussion & Conclusion Guide	references/discussion_conclusion_guide.md	How to write Discussion/Conclusion per Q1 standards
IMRAD Structure	references/imrad_structure.md	Paper structure, section-by-section guidance
Writing Principles	references/writing_principles.md	Clarity, conciseness, accuracy
Figures & Tables	references/figures_tables.md	When and how to use tables vs figures
Paper Template	assets/paper_template.md	Full paper structure template

Read references/discussion_conclusion_guide.md before writing Discussion and Conclusion.

Workflow

Phase 1: Read All Inputs

1. Read every input file listed above
2. Verify all exist — if any are missing, report to user and STOP
3. Copy assets/paper_template.md as the starting structure

Phase 2: Assemble Introduction

Source: idea01.md

1. Read idea01.md — extract the Candidate Paper Titles (2-3 options)
2. Present titles to user: "Which title would you like for the paper?" Wait for user selection.
3. Use selected title as # [Paper Title] at the top of the document
4. Read idea01.md Sections 1-3 (Background, Research Question, Innovation Points)
5. Rewrite as flowing prose (no bullet points in main text)
6. Structure as: Background → Problem Statement → Contributions → Paper Organization
7. Apply Introduction writing pattern per references/watermark_protocol.md Method 2
8. Add [Figure 1: Research overview] placeholder
9. Generate figure prompt: output/figure_prompts/paper/fig_introduction_overview.md

Phase 3: Integrate Literature Review

Source: literature_review.md + review_tables/*.csv

1. Read literature_review.md completely
2. Integrate as Section 2, renumber subsections
3. Embed ALL review summary tables from CSV files directly into the text (as markdown tables)
4. Keep figure placeholders from skill-5 (PRISMA, taxonomy, etc.)
5. Ensure all citations are preserved with original numbering (will renumber later)

Phase 4: Integrate Methodology

Source: methodology.md + ontology03.md

1. Read methodology.md from SubAgent2
2. Integrate as Section 3
3. Embed key tables from ontology03.md:
   • Entity types table
   • Relationship types table
   • Data source summary table
4. Keep figure placeholders (system architecture, KG pipeline, retrieval architecture)
5. Rewrite any bullet points as flowing prose

Phase 5: Integrate Results

Source: resultreport.md + evaluation.md + evaluation CSV tables

1. Read both reports
2. Combine into Section 4, organized by sub-topic:
   • 4.1 KG Statistics (from resultreport.md)
   • 4.2 Extraction Quality (from eval1 CSVs)
   • 4.3 Retrieval Comparison (from eval3 CSVs)
   • 4.4 QA Quality (from eval4 CSVs)
   • 4.5 LLM Comparison (from eval5 CSVs)
   • 4.6 Chatbot Demo (from resultreport.md)
3. Embed tables from CSV files — these are the actual data tables:
   • entity_scores_table.csv → markdown table in 4.2
   • retrieval_comparison_table.csv → markdown table in 4.3
   • qa_metrics_table.csv → markdown table in 4.4
   • llm_comparison_table.csv → markdown table in 4.5
4. Present results objectively — no interpretation here (save for Discussion)
5. Keep figure placeholders

Phase 6: Write Discussion (NEW CONTENT)

This is the most important section skill-6 writes. Read references/discussion_conclusion_guide.md first.

Structure:

1. 5.1 Key Findings (1 paragraph)

   • Restate main findings without repeating numbers
   • State whether research questions were answered

2. 5.2 Interpretation and Insights (3-4 paragraphs)

   • For EACH major result from Section 4:
     • What does it mean?
     • How does it compare to existing work? (cite review papers)
     • Why did this happen? (propose explanations)
     • What is the non-obvious insight?
   • Use WebSearch to find 3-5 recent papers to compare against
   • Provide genuine insights, not just result paraphrasing

3. 5.3 Practical Implications (1-2 paragraphs)

   • What does this mean for practitioners in the domain?
   • How could this system be deployed or adapted?

4. 5.4 Strengths and Limitations (1 paragraph)

   • Be specific, not generic
   • Connect limitations to their impact on results

5. 5.5 Future Research Directions (1 paragraph)

   • Specific, actionable, connected to limitations
   • 3-4 concrete suggestions

Phase 7: Write Conclusion (NEW CONTENT)

Read references/discussion_conclusion_guide.md Part 2.

Structure (3-4 paragraphs):

1. Research summary: Objective + approach + key finding (3-4 sentences)
2. Contributions: Numbered list of specific contributions (4-6 items)
3. Limitations: Brief, honest (2-3 sentences)
4. Future outlook: Forward-looking closing statement (2-3 sentences)

Total: ~400-600 words. Concise. No new arguments.

Phase 8: Write Abstract (LAST)

After all sections are complete:

1. Read the entire assembled paper
2. Write a single flowing paragraph, ≤400 words
3. Cover: Context → Approach → Key Results (with numbers) → Significance
4. No citations in abstract
5. No labeled sections (Background:, Methods:, etc.) — flowing prose only

Phase 9: Merge References

1. Collect all citations from:
   • Literature review (skill-5 references)
   • Methodology (technical references)
   • Discussion (comparison papers)
2. Deduplicate by DOI
3. Renumber sequentially [1], [2], [3]...
4. Update all in-text citations to match new numbering
5. Format consistently (Nature/Vancouver style for engineering journals)

Phase 10: Generate Figure Prompts

For any new figures needed (not already created by upstream skills):

1. output/figure_prompts/paper/fig_introduction_overview.md — Research overview diagram
2. Any additional figures identified during writing

Phase 11: Final Assembly & Polish

1. Assemble complete output/papers/paper_draft.md
2. Ensure continuous prose — convert any remaining bullet points
3. Verify all table numbers are sequential (Table 1, 2, 3...)
4. Verify all figure placeholders are numbered (Figure 1, 2, 3...)
5. Check total word count (~6000-8000 words for AEI-style paper)
6. Run scripts/verify_citations.py from skill-5 if available
7. Apply project signature per references/watermark_protocol.md

Phase 12: Update plan04.md

Follow Auto-Update Protocol for Phase H tasks (H.1 through H.7).

Key Rules

1. Single output file: Everything in one paper_draft.md — not scattered across files.
2. Tables embedded, not referenced: Review/results/evaluation tables go INTO the md as markdown tables. Raw data CSVs (e.g., news collection CSVs) do NOT.
3. Discussion adds insight: Not a result summary. Compare with literature, explain why, extract non-obvious conclusions.
4. Conclusion is concise: Contributions + limitations + future. No new arguments.
5. Abstract written last: After everything else is complete. ≤400 words.
6. Flowing prose: No bullet points in Introduction, Results, Discussion, Conclusion. Only allowed in Methods (e.g., criteria lists).
7. Figures as placeholders: [Figure N: description] only. Actual generation by skill-8.
8. References merged: Single numbered reference list from all sources.
9. Read all references before writing: Especially discussion_conclusion_guide.md.
10. Always update plan04.md.

Composing with Other Skills

/idea-checking            → idea01.md (Introduction framework)
/review-writing           → literature_review.md + tables (Section 2)
@kg-graphrag-builder-agent → methodology.md + resultreport.md (Sections 3-4)
@evaluation-agent          → evaluation.md + CSV tables (Section 4)
/paper-writing             → paper_draft.md (full paper)  ← YOU ARE HERE
/paper-doc                 → converts to .docx (skill-7)