Workflow 3: Paper Writing Pipeline

Orchestrate a complete paper writing workflow for: $ARGUMENTS

Overview

This skill chains five sub-skills into a single automated pipeline:

/paper-plan → /paper-figure → /paper-write → /paper-compile → /auto-paper-improvement-loop
  (outline)     (plots)        (LaTeX)        (build PDF)       (review & polish ×2)

Each phase builds on the previous one's output. The final deliverable is a polished, reviewed

paper/

directory with LaTeX source and compiled PDF.

In this hybrid pack, the pipeline itself is unchanged, but

paper-plan

and

paper-write

use Orchestra-adapted shared references for stronger story framing and prose guidance.

Constants

• VENUE =

  ICLR

  — Target venue. Options:

  ICLR

  ,

  NeurIPS

  ,

  ICML

  ,

  CVPR

  ,

  ACL

  ,

  AAAI

  ,

  ACM

  ,

  IEEE_JOURNAL

  (IEEE Transactions / Letters),

  IEEE_CONF

  (IEEE conferences). Affects style file, page limit, citation format.
• MAX_IMPROVEMENT_ROUNDS = 2 — Number of review→fix→recompile rounds in the improvement loop.
• REVIEWER_MODEL =

  gpt-5.4

  — Model used via Codex MCP for plan review, figure review, writing review, and improvement loop.
• AUTO_PROCEED = true — Auto-continue between phases. Set

  false

  to pause and wait for user approval after each phase.
• HUMAN_CHECKPOINT = false — When

  true

  , the improvement loop (Phase 5) pauses after each round's review to let you see the score and provide custom modification instructions. When

  false

  (default), the loop runs fully autonomously. Passed through to

  /auto-paper-improvement-loop

  .
• ILLUSTRATION =

  figurespec

  — Architecture/illustration generator for Phase 2b:

  figurespec

  (default, deterministic JSON→SVG via

  /figure-spec

  , best for architecture/workflow/topology),

  gemini

  (AI-generated via

  /paper-illustration

  , best for qualitative method illustrations; needs

  GEMINI_API_KEY

  ),

  mermaid

  (Mermaid syntax via

  /mermaid-diagram

  , free, best for flowcharts), or

  false

  (skip Phase 2b, manual only).

Override inline:

/paper-writing "NARRATIVE_REPORT.md" — venue: NeurIPS, illustration: gemini, human checkpoint: true

IEEE example:

/paper-writing "NARRATIVE_REPORT.md" — venue: IEEE_JOURNAL

Inputs

This pipeline accepts one of:

1. NARRATIVE_REPORT.md

   (best) — structured research narrative with claims, experiments, results, figures
2. Research direction + experiment results — the skill will help draft the narrative first
3. Existing

   PAPER_PLAN.md

   — skip Phase 1, start from Phase 2

The more detailed the input (especially figure descriptions and quantitative results), the better the output.

Pipeline

Phase 1: Paper Plan

Invoke

/paper-plan

to create the structural outline:

/paper-plan "$ARGUMENTS"

What this does:

• Parse NARRATIVE_REPORT.md for claims, evidence, and figure descriptions
• Build a Claims-Evidence Matrix — every claim maps to evidence, every experiment supports a claim
• Design section structure (5-8 sections depending on paper type)
• Plan figure/table placement with data sources
• Scaffold citation structure
• GPT-5.4 reviews the plan for completeness

Output:

PAPER_PLAN.md

with section plan, figure plan, citation scaffolding.

Checkpoint: Present the plan summary to the user.

📐 Paper plan complete:
- Title: [proposed title]
- Sections: [N] ([list])
- Figures: [N] auto-generated + [M] manual
- Target: [VENUE], [PAGE_LIMIT] pages

Shall I proceed with figure generation?

• User approves (or AUTO_PROCEED=true) → proceed to Phase 2.
• User requests changes → adjust plan and re-present.

Phase 2: Figure Generation

Invoke

/paper-figure

to generate data-driven plots and tables:

/paper-figure "PAPER_PLAN.md"

What this does:

• Read figure plan from PAPER_PLAN.md
• Generate matplotlib/seaborn plots from JSON/CSV data
• Generate LaTeX comparison tables
• Create

  figures/latex_includes.tex

  for easy insertion
• GPT-5.4 reviews figure quality and captions

Output:

figures/

directory with PDFs, generation scripts, and LaTeX snippets.

Scope:

paper-figure

covers data plots and comparison tables. Architecture diagrams, pipeline figures, and method illustrations are handled in Phase 2b below.

Phase 2b: Architecture & Illustration Generation

Skip this step entirely if

illustration: false

.

If the paper plan includes architecture diagrams, pipeline figures, audit cascades, or method illustrations, invoke the appropriate generator based on the

illustration

parameter:

When

illustration: figurespec

(default) — invoke

/figure-spec

:

/figure-spec "[architecture/workflow description from PAPER_PLAN.md]"

• Deterministic JSON → SVG vector rendering (editable, reproducible)
• Best for: system architecture, workflow pipelines, audit cascades, layered topology
• Output:

  figures/*.svg

  +

  figures/*.pdf

  (via rsvg-convert) +

  figures/specs/*.json

• No external API, runs fully local

When

illustration: gemini

— invoke

/paper-illustration

:

/paper-illustration "[method description from PAPER_PLAN.md or NARRATIVE_REPORT.md]"

• Claude plans → Gemini optimizes → Nano Banana Pro renders → Claude reviews (score ≥ 9)
• Best for: qualitative method illustrations, natural-style diagrams, result grids
• Output:

  figures/ai_generated/*.png

• Requires

  GEMINI_API_KEY

  environment variable

When

illustration: mermaid

— invoke

/mermaid-diagram

:

/mermaid-diagram "[method description from PAPER_PLAN.md]"

• Generates Mermaid syntax diagrams (flowchart, sequence, class, state, etc.)
• Best for: lightweight flowcharts, state machines, simple sequence diagrams
• Output:

  figures/*.mmd

  +

  figures/*.png

• Free, no API key needed

When

illustration: false

— skip entirely. All non-data figures must be created manually (draw.io, Figma, TikZ) and placed in

figures/

before Phase 3.

Choosing the right mode:

• Formal architecture / workflow / topology figures →

  figurespec

  (default)
• Method concept illustrations with natural style →

  gemini

• Quick flowchart / state machine →

  mermaid

• Full manual control →

  false

These are complementary, not mutually exclusive: you can run multiple generators for different figures in the same paper by re-invoking with different

illustration

overrides.

Checkpoint: List generated vs manual figures.

📊 Figures complete:
- Data plots (auto, Phase 2): [list]
- Architecture/illustrations (auto, Phase 2b, mode=<illustration>): [list]
- Manual (need your input): [list]
- LaTeX snippets: figures/latex_includes.tex

[If manual figures needed]: Please add them to figures/ before I proceed.
[If all auto]: Shall I proceed with LaTeX writing?

Phase 3: LaTeX Writing

Invoke

/paper-write

to generate section-by-section LaTeX:

/paper-write "PAPER_PLAN.md"

What this does:

• Write each section following the plan, with proper LaTeX formatting
• Insert figure/table references from

  figures/latex_includes.tex

• Build

  references.bib

  from citation scaffolding
• Clean stale files from previous section structures
• Automated bib cleaning (remove uncited entries)
• De-AI polish (remove "delve", "pivotal", "landscape"...)
• GPT-5.4 reviews each section for quality

Output:

paper/

directory with

main.tex

,

sections/*.tex

,

references.bib

,

math_commands.tex

.

Checkpoint: Report section completion.

✍️ LaTeX writing complete:
- Sections: [N] written ([list])
- Citations: [N] unique keys in references.bib
- Stale files cleaned: [list, if any]

Shall I proceed with compilation?

Phase 4: Compilation

Invoke

/paper-compile

to build the PDF:

/paper-compile "paper/"

What this does:

• latexmk -pdf

  with automatic multi-pass compilation
• Auto-fix common errors (missing packages, undefined refs, BibTeX syntax)
• Up to 3 compilation attempts
• Post-compilation checks: undefined refs, page count, font embedding
• Precise page verification via

  pdftotext

• Stale file detection

Output:

paper/main.pdf

Checkpoint: Report compilation results.

🔨 Compilation complete:
- Status: SUCCESS
- Pages: [X] (main body) + [Y] (references) + [Z] (appendix)
- Within page limit: YES/NO
- Undefined references: 0
- Undefined citations: 0

Shall I proceed with the improvement loop?

Phase 4.5: Proof Verification (theory papers only)

Skip this phase if the paper contains no theorems, lemmas, or proofs.

if paper contains \begin{theorem} or \begin{lemma} or \begin{proof}:
    Run /proof-checker "paper/"
    This invokes GPT-5.4 xhigh to:
    - Verify all proof steps (hypothesis discharge, interchange justification, etc.)
    - Check for logic gaps, quantifier errors, missing domination conditions
    - Attempt counterexamples on key lemmas
    - Generate PROOF_AUDIT.md with issue list + severity

    If FATAL or CRITICAL issues found:
        Fix before proceeding to improvement loop
    If only MAJOR/MINOR:
        Proceed, improvement loop may address remaining issues
else:
    skip — no proofs, no action

Phase 4.7: Paper Claim Audit

Skip if no result files exist (e.g., survey/position papers with no experiments).

if results/*.json or results/*.csv or outputs/*.json exist:
    Run /paper-claim-audit "paper/"
    Fresh zero-context reviewer compares every number in the paper
    against raw result files. Catches rounding inflation, best-seed
    cherry-pick, config mismatch, delta errors.

    If FAIL:
        Fix mismatched numbers before improvement loop
    If WARN:
        Proceed, but flag for manual verification
else:
    skip — no experimental results to verify

Phase 5: Auto Improvement Loop

Invoke

/auto-paper-improvement-loop

to polish the paper:

/auto-paper-improvement-loop "paper/"

What this does (2 rounds):

Round 1: GPT-5.4 xhigh reviews the full paper → identifies CRITICAL/MAJOR/MINOR issues → Claude Code implements fixes → recompile → save

main_round1.pdf

Round 2: GPT-5.4 xhigh re-reviews with conversation context → identifies remaining issues → Claude Code implements fixes → recompile → save

main_round2.pdf

Typical improvements:

• Fix assumption-model mismatches
• Soften overclaims to match evidence
• Add missing interpretations and notation
• Strengthen limitations section
• Add theory-aligned experiments if needed

Output: Three PDFs for comparison +

PAPER_IMPROVEMENT_LOG.md

.

Format check (included in improvement loop Step 8): After final recompilation, auto-detect and fix overfull hboxes (content exceeding margins), verify page count vs venue limit, and ensure compact formatting. Location-aware thresholds: any main-body overfull blocks completion regardless of size; appendix overfulls block only if >10pt; bibliography overfulls block only if >20pt.

Phase 5.5: Final Paper Claim Audit (MANDATORY submission gate)

After

/auto-paper-improvement-loop

finishes, rerun

/paper-claim-audit

before the final report whenever the paper contains numeric claims and machine-readable raw result files exist.

Use the same detectors as Phase 4.7:

• numeric-claim regex over

  paper/main.tex

  and

  paper/sections/*.tex

• raw-evidence file search in

  results/

  ,

  outputs/

  ,

  experiments/

  , and

  figures/

  for

  .json

  ,

  .jsonl

  ,

  .csv

  ,

  .tsv

  ,

  .yaml

  , or

  .yml

This phase is mandatory if both detectors are positive. It blocks the final report. If numeric claims exist but no raw result files are found, stop and warn the user before declaring the paper complete. If no numeric claims exist, skip.

NUMERIC_CLAIMS=$(rg -n -e '[0-9]+(\.[0-9]+)?\s*(%|\\%|±|\\pm|x|×)' \
  -e '(accuracy|BLEU|F1|AUC|mAP|top-1|top-5|error|loss|perplexity|speedup|improvement)' \
  paper/main.tex paper/sections 2>/dev/null || true)

RAW_RESULT_FILES=$(find results outputs experiments figures -type f \
  \( -name '*.json' -o -name '*.jsonl' -o -name '*.csv' -o -name '*.tsv' -o -name '*.yaml' -o -name '*.yml' \) 2>/dev/null | head -200)

if [ -n "$NUMERIC_CLAIMS" ] && [ -n "$RAW_RESULT_FILES" ]; then
    Run /paper-claim-audit "paper/"
    If FAIL:
        Fix mismatched numbers before the final report
elif [ -n "$NUMERIC_CLAIMS" ]; then
    Stop and warn: the paper contains numeric claims but no raw evidence files were found
fi

Empirical motivation: in our April 2026 NeurIPS run, the final paper claimed

w ∈ {0,1,2,3}

for the width-tradeoff experiment but the raw JSON had

w ∈ {0,1,2,3,4,5}

. The crossing-point tolerance was claimed as

0.05%

but the actual relative error was

0.0577%

. Both were caught only after manual

paper-claim-audit

invocation in the final round; the improvement loop did not detect them.

Phase 5.8: Citation Audit (submission gate)

After the final paper-claim-audit passes, run

/citation-audit

to verify every

\cite{...}

along three axes: existence, metadata correctness, and context appropriateness. This is the fourth and final layer of the evidence-and-claim assurance stack (

experiment-audit

→

result-to-claim

→

paper-claim-audit

→

citation-audit

).

if paper/references.bib (or paper.bib) exists and contains entries cited from sec/*.tex:
    Run /citation-audit "paper/"
    Fresh cross-family reviewer (gpt-5.4 via Codex MCP) with web/DBLP/arXiv lookup
    verifies each entry:
      (i)   EXISTENCE — paper resolves at claimed arXiv ID / DOI / venue
      (ii)  METADATA — author names, year, venue, title match canonical sources
      (iii) CONTEXT — cited paper actually establishes the claim it supports

    Output:
      - CITATION_AUDIT.md (human-readable per-entry verdict report)
      - CITATION_AUDIT.json (machine-readable verdict ledger)
      - Per-entry verdicts: KEEP / FIX / REPLACE / REMOVE

    If any REPLACE or REMOVE verdicts:
        Surface to user for human approval — never auto-modify content claims
    If only FIX verdicts (metadata corrections):
        Apply with user confirmation, then recompile
    If all KEEP:
        Pass — bibliography clean for submission
else:
    skip — no bib file or no citations

Why this is the most diagnostic of the four audit layers: wildly fake citations are easy to spot. The dangerous failure mode is a real paper used to support a claim it does not actually establish (wrong-context citations) — these slip past metadata-only checks and damage submission credibility. Run cost is wall-clock heavy (web lookup per entry); run once per submission, not per save.

Empirical motivation: in our April 2026 ARIS technical-report run, two real papers (

madaan2023selfrefine

,

liu2023reviewergpt

) were cited in contexts they did not actually support, and one entry (

hidden2025aiscientistpitfalls

) had

author = "Anonymous"

because the metadata had not been resolved. None were caught by the improvement loop or numeric claim audit; only fresh web-lookup review surfaced them.

Phase 6: Final Report

# Paper Writing Pipeline Report

**Input**: [NARRATIVE_REPORT.md or topic]
**Venue**: [ICLR/NeurIPS/ICML/CVPR/ACL/AAAI/ACM/IEEE_JOURNAL/IEEE_CONF]
**Date**: [today]

## Pipeline Summary

| Phase | Status | Output |
|-------|--------|--------|
| 1. Paper Plan | ✅ | PAPER_PLAN.md |
| 2. Figures | ✅ | figures/ ([N] auto + [M] manual) |
| 3. LaTeX Writing | ✅ | paper/sections/*.tex ([N] sections, [M] citations) |
| 4. Compilation | ✅ | paper/main.pdf ([X] pages) |
| 5. Improvement | ✅ | [score0]/10 → [score2]/10 |
| 5.5 Final Claim Audit | ✅/SKIP | PAPER_CLAIM_AUDIT.{md,json} |
| 5.8 Citation Audit | ✅/SKIP | CITATION_AUDIT.{md,json} |

## Improvement Scores
| Round | Score | Key Changes |
|-------|-------|-------------|
| Round 0 | X/10 | Baseline |
| Round 1 | Y/10 | [summary] |
| Round 2 | Z/10 | [summary] |

## Deliverables
- paper/main.pdf — Final polished paper
- paper/main_round0_original.pdf — Before improvement
- paper/main_round1.pdf — After round 1
- paper/main_round2.pdf — After round 2
- paper/PAPER_IMPROVEMENT_LOG.md — Full review log
- paper/PAPER_CLAIM_AUDIT.{md,json} — Numerical claim verification (if Phase 5.5 ran)
- paper/CITATION_AUDIT.{md,json} — Bibliography verification (if Phase 5.8 ran)

## Remaining Issues (if any)
- [items from final review that weren't addressed]

## Next Steps
- [ ] Visual inspection of PDF
- [ ] Add any missing manual figures
- [ ] Submit to [venue] via OpenReview / CMT / HotCRP

Output Protocols

Follow these shared protocols for all output files:

• Output Versioning Protocol — write timestamped file first, then copy to fixed name
• Output Manifest Protocol — log every output to MANIFEST.md
• Output Language Protocol — note: paper-writing always outputs English LaTeX for venue submission

Key Rules

• Large file handling: If the Write tool fails due to file size, immediately retry using Bash (

  cat << 'EOF' > file

  ) to write in chunks. Do NOT ask the user for permission — just do it silently.
• Don't skip phases. Each phase builds on the previous one — skipping leads to errors.
• Checkpoint between phases when AUTO_PROCEED=false. Present results and wait for approval.
• Manual figures first. If the paper needs architecture diagrams or qualitative results, the user must provide them before Phase 3.
• Compilation must succeed before entering the improvement loop. Fix all errors first.
• Preserve all PDFs. The user needs round0/round1/round2 for comparison.
• Document everything. The pipeline report should be self-contained.
• Respect page limits. If the paper exceeds the venue limit, suggest specific cuts before the improvement loop.

Composing with Other Workflows

/idea-discovery "direction"         ← Workflow 1: find ideas
implement                           ← write code
/run-experiment                     ← deploy experiments
/auto-review-loop "paper topic"     ← Workflow 2: iterate research
/paper-writing "NARRATIVE_REPORT.md"  ← Workflow 3: you are here
                                         submit! 🎉

Or use /research-pipeline for the Workflow 1+2 end-to-end flow,
then /paper-writing for the final writing step.

Typical Timeline

Phase	Duration	Can sleep?
1. Paper Plan	5-10 min	No
2. Figures	5-15 min	No
3. LaTeX Writing	15-30 min	Yes ✅
4. Compilation	2-5 min	No
5. Improvement	15-30 min	Yes ✅

Total: ~45-90 min for a full paper from narrative report to polished PDF.