paper-write

name	paper-write
description	Draft LaTeX paper section by section from an outline. Use when user says "write paper", "draft LaTeX", or wants to generate LaTeX content from a paper plan.
argument-hint	["venue-or-section"]
allowed-tools	Bash(*), Read, Write, Edit, Grep, Glob, Agent, WebSearch, WebFetch, mcp__codex__codex, mcp__codex__codex-reply

Paper Write: Section-by-Section LaTeX Generation

Draft a LaTeX paper based on: $ARGUMENTS

Constants

• TARGET_VENUE = ICLR — Supported: ICLR, NeurIPS, ICML. Override via Additional Parameters.
• MAX_PAGES = 9 — Main body to Conclusion end. Refs/appendix excluded. Body pages must be ≥ MAX_PAGES.
• ANONYMOUS = true
• DBLP_BIBTEX = true — Fetch real BibTeX from DBLP/CrossRef. Never fabricate.
• CUSTOM_REQUIREMENTS — Highest priority.
• REVIEWER_SCRIPT — External reviewer script

Inputs

1. PAPER_PLAN.md — outline with claims-evidence matrix, figure plan
2. NARRATIVE_REPORT.md — research narrative
3. experiment_results.md — structured experiment results (from experiment-bridge)
4. figures/ — PDFs + latex_includes*.tex + experiment_data.json
5. Existing .bib file (or will create)

If no PAPER_PLAN.md, generate minimal outline from available docs.

Orchestra References (use when needed)

• ../shared-references/writing-principles.md — story framing, clarity
• ../shared-references/venue-checklists.md — submission requirements
• ../shared-references/citation-discipline.md — citation fallback

Load shared rules

cat _utils/writing_rules.md 2>/dev/null || cat skills/shared-scripts/writing_rules.md

Workflow

Step 0: Backup + resume check + upstream validation

⛔ 上游输出完整性检查（写论文前必做）：

echo "=== Upstream outputs validation ==="
UPSTREAM_OK=true

# 1. 核心文件是否存在
for f in PAPER_PLAN.md RESULTS.md; do
    if [ -f "$f" ]; then
        sz=$(wc -c < "$f")
        echo "✅ $f ($sz chars)"
        [ "$sz" -lt 500 ] && { echo "  ⚠ File too small, content may be incomplete"; UPSTREAM_OK=false; }
    else
        echo "⚠ $f not found (paper-write will use minimal outline)"
    fi
done

# 2. 实验数据文件
[ -f figures/all_results.json ] && echo "✅ figures/all_results.json" || echo "⚠ No all_results.json — numerical values may be inaccurate"
[ -f experiment_results.md ] && echo "✅ experiment_results.md" || echo "  (no experiment_results.md, will rely on RESULTS.md)"

# 3. 图表文件
PDF_COUNT=$(ls figures/*.pdf 2>/dev/null | wc -l)
echo "Figures: $PDF_COUNT PDFs"
[ "$PDF_COUNT" -eq 0 ] && echo "⚠ No PDF figures — paper will lack visual content"

# 4. latex_includes.tex 是否存在
[ -f figures/latex_includes.tex ] &&   ||  


 [ -f PAPER_PLAN.md ]; 
    CLAIM_ROWS=$(grep -c  PAPER_PLAN.md 2>/dev/null ||  0)
    [  -gt 2 ] &&   ||  


 
 ||  

Back up existing paper/ to paper-backup-{timestamp}/. Clean stale section files. Check for incomplete sections:

echo "=== Resume check ==="
if [ -d "paper/sections" ]; then
    for f in paper/sections/*.tex; do
        [ -f "$f" ] || continue
        chars=$(wc -c < "$f")
        if [ "$chars" -lt 500 ]; then
            echo "⚠ Placeholder: $(basename $f) ($chars chars) — needs writing"
        else
            echo "✅ Complete: $(basename $f) ($chars chars)"
        fi
    done
fi

Resume: only write placeholder sections (<500 chars or contains "placeholder"/"TODO"), skip completed ones (>2000 chars). See <resume_strategy> in writing_rules.md.

Step 1: Initialize

Create paper/, copy venue template, generate math_commands.tex (paper-specific commands only), create section files.

Step 1.5: Figure inventory

Before writing any section, build a complete inventory of available figures:

echo "=== Available PDF figures ==="
ls -la figures/*.pdf 2>/dev/null || echo "No PDF figures found"
echo ""
echo "=== latex_includes.tex content (figure→PDF mapping) ==="
cat figures/latex_includes.tex 2>/dev/null || echo "No latex_includes.tex"
echo ""
echo "=== TikZ diagrams ==="
[ -s figures/tikz_architecture_examples.tex ] && echo "tikz_architecture_examples.tex exists" || echo "No TikZ diagrams"

⛔ Build a FIGURE EMBEDDING PLAN before writing any section:

FIGURE EMBEDDING PLAN:
1. fig_main_results.pdf → Experiments section
2. fig_ablation.pdf → Experiments section
3. fig_training_curves.pdf → Experiments section
4. TABLE_main.tex → Experiments section
5. tikz_architecture (from tikz_architecture_examples.tex) → Method section

• Must use figure blocks from latex_includes.tex, not write \includegraphics from scratch
• TikZ diagrams must be embedded into corresponding sections
• Read experiment_results.md / RESULTS.md for exact numbers — do not invent results

⛔ CRITICAL: ALL numerical results in the paper MUST come from figures/all_results.json or RESULTS.md. Before writing any results/experiments section, run:

[ -f figures/all_results.json ] && cat figures/all_results.json
[ -f RESULTS.md ] && cat RESULTS.md
[ -f experiment_results.md ] && cat experiment_results.md

When quoting specific numbers (accuracy, RMSE, F1, p-values, speedup ratios, parameter counts, etc.), you MUST copy them verbatim from these files. Do NOT estimate, round, or make up values from LLM memory. A paper with fabricated numbers will fail the final quality gate's numerical consistency check.

⛔ Claims-Evidence 对照（必须严格遵循规划）：

Before writing each section, re-read PAPER_PLAN.md's claims-evidence matrix:

# 提取 PAPER_PLAN.md 中的 claims-evidence 表
grep -A 100 'Claims-Evidence\|claim.*evidence\|claim-evidence' PAPER_PLAN.md 2>/dev/null | head -30

Writing discipline:

• Every claim in the paper MUST trace back to a row in the matrix
• Do not add new claims not in the plan (if you discover something, update PAPER_PLAN.md first)
• Do not skip claims that were planned (even negative results should be reported)
• Each claim's numerical evidence must match the value in figures/all_results.json

If a planned claim has no evidence in the data, write an honest statement like "preliminary results suggest X, though we leave formal validation to future work" instead of fabricating evidence.

Step 1.5: Pre-fetch verified reference pool (BEFORE writing any text)

⛔ This step MUST happen before Step 2. Do NOT write any \citep{} until this pool exists.

The goal is to build a pool of real, verified papers so that when writing body text, you only cite papers that actually exist.

PYTHON=$(command -v python3 2>/dev/null || command -v python 2>/dev/null)
mkdir -p _tmp

# Search for papers in each key topic area of this paper
# (adapt these queries to your specific paper topic)
echo "=== Searching key topic areas ==="

# Extract topic keywords from PAPER_PLAN.md
grep -i 'related\|background\|literature\|baseline\|prior work' PAPER_PLAN.md 2>/dev/null | head -20

# For each major topic/method mentioned in the plan, search for real papers:
# Example queries (REPLACE with your actual topics):
#   $PYTHON "$SCHOLAR_SCRIPT" bibtex "spatial Durbin model digital economy" --max 5
#   $PYTHON "$SCHOLAR_SCRIPT" bibtex "computing infrastructure regional development" --max 5
#   $PYTHON "$SCHOLAR_SCRIPT" bibtex "spatial spillover effect panel data" --max 5

After searching, create _tmp/_verified_refs.txt with one line per verified paper:

key: lesage_2009_spatial_econometrics | title: Introduction to Spatial Econometrics | authors: LeSage, Pace | year: 2009 | match: good
key: elhorst_2014_spatial_panel | title: Spatial Econometrics: From Cross-Sectional Data to Spatial Panels | authors: Elhorst | year: 2014 | match: good

When writing body text in Step 2, ONLY use citation keys from this verified pool. If you need to cite a paper not in the pool, search for it first and add it to the pool before citing.

Fallback: If scholar_fetch.py returns no results or match_label="low" for a topic, use WebSearch to find the paper on Google Scholar / Semantic Scholar website, then manually verify title + authors + year before adding to the pool.

Step 2: Write each section

Writing order: Method → Experiments → Introduction → Related Work → Conclusion (core content first). Save each section immediately. If approaching output limit, create % [PLACEHOLDER] files.

⛔ Writing style rules:

• No \begin{itemize} or \begin{enumerate} in body text — bullet lists are the #1 AI writing tell. Use flowing prose with inline numbering "(1)...(2)...(3)..." or transition words "First,...Second,...Finally,...".
• Each paragraph must have ≥3 sentences. No 1-2 sentence micro-paragraphs.
• Consecutive paragraphs must not start with the same phrase.

Follow all rules from _utils/writing_rules.md (interleaving, embedding, LaTeX constraints).

For each section, copy the matching figure/table blocks from figures/latex_includes.tex (or figures/*.tex) into the section file. Path: always ../figures/xxx.pdf (relative to paper/). Use [H] float specifier. Post-write check: every \ref must have matching \label.

Wide tables (≥6 columns or multiple p{} columns): wrap with \resizebox{\textwidth}{!}{...}.

After each section, check chars:

chars=$(wc -c < "paper/sections/current_section.tex")
echo "Current section: $chars chars"
# English LaTeX ≈ 2000-2500 chars/page
# If section page budget is 2 pages but only 2000 chars (~1 page), expand immediately

<exemplar_depth>

Writing depth by venue

ICLR/NeurIPS/ICML (9 pages main body):

• Abstract (0.3p): what → why hard → how → evidence → strongest result. 150-250 words. Self-contained
• Introduction (1.5p): hook → gap → contributions → results preview → hero figure. Front-load the contribution
• Related Work (1-1.5p): organize by category, synthesize not list. Each category: 3-5 papers with method summary + positioning vs this work
• Method (2-2.5p): notation → formulation → algorithm. Every formula has intuition explanation. Key derivation steps not skipped
• Experiments (3-4p): setup → main results table → comparison plots → ablation table → analysis. Every result has 1-2 paragraphs of interpretation (not just "our method outperforms")
• Conclusion (0.5p): rephrase contributions + limitations + future work

JMLR/TPAMI journal (15-20 pages):

• Introduction (2-3p): more thorough literature positioning
• Related Work (2-3p): comprehensive survey by sub-topic
• Method (4-6p): full derivations, proofs, complexity analysis
• Experiments (6-8p): multiple datasets, extensive ablations, qualitative analysis, failure cases
• Conclusion (1p): detailed limitations and future directions </exemplar_depth>

Expansion strategies (not padding — substantive content):

• Formula listed without derivation → add step-by-step derivation with intuition
• Result only says "as shown in Table X" → add 1-2 paragraphs of interpretation (what numbers mean, comparison, reasoning)
• Related work only lists papers → add method summaries and positioning vs this work
• Algorithm only has pseudocode → add explanation of key steps and complexity analysis

Section guidelines

• Abstract: what → why hard → how → evidence → strongest result. Self-contained. 150-250 words.
• Introduction: hook → gap → contributions → results preview → hero figure. 1.5 pages. Front-load contribution.
• Related Work: ≥1 full page. Organize by category, synthesize not list.
• Method: notation → formulation → algorithm. 1.5-2 pages.
• Experiments: setup → main results → ablations. 2.5-3 pages. Every claim needs evidence.
• Conclusion: rephrase contributions + limitations + future work. 0.5 pages.

Step 3: Build bibliography

Follow the <references_workflow> in _utils/writing_rules.md. Venue style: natbib (citep/citet). Verify references.bib is non-empty before proceeding.

⛔ Use the scholar_fetch.py tool for ALL reference retrieval. NEVER fabricate BibTeX from memory.

⛔ 引用写法规则：写正文时，citation key 必须包含描述性关键词，格式为 作者姓_年份_主题关键词。 例如：\citep{wang_2023_supply_chain_resilience} 而不是 \citep{wang2023supply}。 这样 Step 3b 搜索时能用关键词找到正确的论文。如果不确定作者/年份，用 TODO__ 前缀：\citep{TODO__digital_economy_spatial_spillover}。

# Step 3a: Collect all cited keys and extract search queries
grep -roh '\\cite[tp]*{[^}]*}' paper/sections/*.tex paper/main.tex 2>/dev/null \
  | grep -oP '\{[^}]+\}' | tr -d '{}' | tr ',' '\n' | sed 's/^ *//;s/ *$//' | sort -u > _tmp/_cited_keys.txt
echo "Cited keys: $(wc -l < _tmp/_cited_keys.txt)"
cat _tmp/_cited_keys.txt

# Step 3b: For each cited key, extract descriptive keywords and search
PYTHON=$(command -v python3 2>/dev/null || command -v python 2>/dev/null)
while IFS= read -r key; do
    # Convert citation key to search query: replace _ with spaces, remove TODO prefix
    query=$(echo "$key" | sed 's/^TODO__//; s/_/ /g')
    echo "--- Fetching: $key (query: $query) ---"
    $PYTHON "$SCHOLAR_SCRIPT" bibtex "$query" --max 3
    sleep 0.5
done < _tmp/_cited_keys.txt

For each result:

1. Check match_label: if "good" → use directly. If "partial" → verify title matches your intent. If "low" → this is likely the wrong paper, search again with better keywords or use WebSearch.
2. Check match_score: score < 0.3 means the search result probably doesn't match what you cited. Do NOT blindly use it.
3. Pick the correct paper and copy its bibtex field into paper/references.bib.
4. Replace the citation key in .tex files with the actual key from the BibTeX entry.
5. If bibtex_source=auto, add % [VERIFY] above the entry.
6. If match_label="low" and no better result found, add % [LOW_MATCH - verify this is the intended paper] and use WebSearch as fallback.

Step 4: De-AI polish

See <de_ai_polish> in _utils/writing_rules.md.

Step 5: Cross-review

Send draft to external reviewer for feedback before finalizing:

mkdir -p _tmp
cat << 'REVIEW_EOF' > _tmp/_review_prompt.txt
Please review this academic paper draft. Focus on:
1. Logical flow and argument structure
2. Claim-evidence alignment (every claim has supporting data?)
3. Writing clarity and conciseness
4. Missing content or weak sections
5. Score (1-10) and top 3 actionable improvements

## Paper sections:
REVIEW_EOF
for f in paper/sections/*.tex; do
    [ -f "$f" ] && echo "### $(basename $f)" >> _tmp/_review_prompt.txt && cat "$f" >> _tmp/_review_prompt.txt
done
PYTHON=$(command -v python3 2>/dev/null || command -v python 2>/dev/null)
$PYTHON "$REVIEWER_SCRIPT" --prompt-file _tmp/_review_prompt.txt --thread-file _tmp/_reviewer_thread.json 2>&1 | tee _tmp/_cross_review.txt

If reviewer script unavailable, skip this step.

Step 6: Reverse outline test

Extract topic sentences → read in sequence → check claim coverage → fix gaps.

Step 7: Final checks

bash _utils/writing_check.sh paper/ 2>/dev/null || bash skills/shared-scripts/writing_check.sh paper/

Figure embedding verification (must pass before finishing):

echo "=== Figure embedding check ==="
missing=0
# Check every PDF in figures/ is referenced in sections
for pdf in figures/*.pdf; do
    [ -f "$pdf" ] || continue
    bn=$(basename "$pdf")
    if ! grep -rq "$bn" paper/sections/*.tex paper/main.tex 2>/dev/null; then
        echo "MISSING: $bn not embedded in any section"
        missing=$((missing + 1))
    fi
done
# Check every label in figures/*.tex is in sections
for fig_tex in figures/*.tex; do
    [ -f "$fig_tex" ] || continue
    for lbl in $(grep -oh '\\label{[^}]*}' "$fig_tex" 2>/dev/null); do
        if ! grep -rq "$lbl" paper/sections/*.tex paper/main.tex 2>/dev/null; then
            echo "MISSING: $lbl (from $(basename $fig_tex)) not in any section"
            missing=$((missing + 1))
        fi
    done
done
echo "Total missing: $missing"

If any figures are missing, go back and embed them into the appropriate sections before finishing. ⛔ Do NOT finish until missing = 0.

Page estimate check:

echo "=== Section sizes ==="
total=0
for f in paper/sections/*.tex; do
    chars=$(wc -c < "$f")
    total=$((total + chars))
    echo "  $(basename $f): $chars chars"
done
echo "  Total: $total chars (~$((total / 2200)) pages), Target: ≥ MAX_PAGES pages"

If total chars < MAX_PAGES × 2000, expand the thinnest sections before finishing.

Key Rules

• Large files: Bash heredoc
• No author info — anonymous block
• Complete sections, not outlines
• One file per section
• Every claim cites evidence
• Venue style: natbib (citep/citet)
• Clean bib — only cited entries
• Section count flexible (5-8)
• Backup before overwrite
• Front-load the contribution
• Primary output: paper/ directory, temp files: _tmp/