Stat Paper Writing

Statistics Paper Writing Pipeline

Orchestrate a complete statistics, applied statistics, or ML theory paper writing workflow for: $ARGUMENTS

Overview

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

/stat-paper-plan → /paper-figure → /stat-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 paper/ directory with LaTeX source, supplement, and compiled PDF.

The improvement loop uses /stat-polishing standards internally to enforce Big Four-style writing (JASA, AoS, JRSS-B, Biometrika): punctuation discipline, AI-template removal, COPSS-style scholar voice, figure design rules, and main-supplement separation. After the pipeline completes, the user can also invoke /stat-polishing directly on specific sections that need further refinement.

Constants

• VENUE = AOS — Target venue. Options:
• Theory/methodology: AOS, JASA, JRSSB, BIOMETRIKA, BERNOULLI, EJS, STATSINICA, MSL
• Application: AOAS, JASA_ACS (alias JASA_APP), BIOSTATISTICS, STATMED, JCGS, JABES
• ML theory conferences: COLT, ALT
• PAPER_TYPE = auto — theory, methodology, application, or auto.
• MAXIMPROVEMENTROUNDS = 2 — Number of review→fix→recompile rounds.
• CLAUDEREVIEWERMODEL = claude-opus-4-6 — Claude subagent model for fast internal reviews in each phase.
• CODEXREVIEWERMODEL = gpt-5.5 — External LLM for Codex MCP reviews at model_reasoning_effort: xhigh.
• REVIEW_MODE = both — Options: claude (fast), codex (deep), both (Claude every round, Codex on final round). Default both. Passed through to plan/write/polish sub-skills.
• AUTO_PROCEED = true — Auto-continue between phases. Set false to pause after each phase.
• HUMAN_CHECKPOINT = false — When true, improvement loop pauses after each round's review.

> Theory example: /stat-paper-writing "NARRATIVE_REPORT.md" — venue: AOS, paper type: theory > Methodology example: /stat-paper-writing "NARRATIVE_REPORT.md" — venue: JASA, paper type: methodology, human checkpoint: true > Application example (AOAS): /stat-paper-writing "APPLICATION_REPORT.md" — venue: AOAS, paper type: application > Application example (JASA ACS): /stat-paper-writing "APPLICATION_REPORT.md" — venue: JASA_ACS, paper type: application > COLT example: /stat-paper-writing "NARRATIVE_REPORT.md" — venue: COLT

Inputs

This pipeline accepts one of:

1. NARRATIVE_REPORT.md (best for theory/methodology) — research narrative with theorems, proofs, simulations, results
2. APPLICATION_REPORT.md + DATA_DESCRIPTION.md (best for application papers) — analysis narrative with dataset description, EDA, scientific findings, validation
3. Theorem statements + simulation results — the skill will help structure a theory/methodology paper
4. Dataset + analysis pipeline + scientific findings — the skill will help structure an application paper
5. Existing PAPER_PLAN.md — skip Phase 1, start from Phase 2

The more detailed the input, the better the output:

• For theory/methodology: theorem statements, proof outlines, simulation designs, quantitative results
• For application papers: dataset details (source, size, variables, time period), EDA findings, statistical challenges, comparison methods used by domain practitioners, substantive scientific findings, validation strategy, domain interpretation

Pipeline

Phase 1: Paper Plan

Invoke /stat-paper-plan to create the structural outline:

/stat-paper-plan "$ARGUMENTS"

What this does:

• Parse input for theorems, claims, evidence, and assumptions
• For application papers: also parse dataset details, scientific question, statistical challenges, findings, validation strategy
• Build matrices:
• Theorems-Evidence Matrix and Claims-Evidence Matrix (all paper types)
• Findings-Evidence Matrix and Data-Challenges Matrix (application papers)
• Determine paper type (theory / methodology / application)
• Design section structure (5-9 sections, with application papers using the 7-section data-first layout)
• Plan assumption organization (lighter for application papers)
• Plan simulation studies (informed by real data for application papers)
• Plan figures and tables (EDA figures emphasized for application papers)
• Scaffold citations (statistical + domain literature for application papers)
• Claude subagent reviews for completeness

Output: PAPER_PLAN.md with section plan, matrices, figure plan.

Checkpoint (theory/methodology):

Paper plan complete:
- Title: [proposed title]
- Type: [theory/methodology]
- Venue: [venue]
- Main theorems: [N]
- Assumptions: [N] ((A1)-(AN))
- Sections: [N] ([list])
- Figures: [N] auto + [M] manual
- Simulations: [N] DGPs × [M] methods

Shall I proceed with figure generation?

Checkpoint (application paper):

Application paper plan complete:
- Title: [proposed title]
- Type: application
- Venue: [AOAS/JASA_ACS/Biostatistics/etc.]
- Scientific question: [one-line summary]
- Dataset: [name, size, time period]
- Statistical challenges identified: [N]
- Main theorems (light): [1-2 max]
- Sections: [N] ([list, with §6 Application as centerpiece])
- EDA figures planned: [N]
- Application figures planned: [N]
- Simulation DGPs (real-data-informed): [N]
- Comparison methods (incl. domain-standard): [N]
- Substantive findings preview: [N findings]

Shall I proceed with figure generation?

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/ggplot plots:
• For theory/methodology papers:
• Convergence rate verification (log-log plots)
• Method comparison (box plots, tables)
• Coverage probability plots
• Power curves
• Rate comparison tables
• For application papers (additionally):
• EDA figures (data distributions, dependence structure, missingness, time series)
• Descriptive statistics tables
• Application main analysis figures (model fits, parameter estimates with CI)
• Comparison with domain-standard methods
• Validation plots (holdout, calibration, CV)
• Sensitivity analysis figures
• Generate LaTeX comparison tables
• Create figures/latex_includes.tex (and latex_includes_eda.tex for application papers)

Statistics-specific figure requirements:

• Log-log plots for rate verification must include theoretical rate line
• Simulation tables must include standard errors
• Confidence interval coverage plots should include nominal level line
• Box plots or violin plots for distribution comparison across methods

Application-specific figure requirements:

• EDA figures must reveal the statistical challenges discussed in §2.4
• Every figure in §6 (Application) must have a self-contained caption
• Multi-panel figures should walk the reader through the analysis
• Use colorblind-safe palettes
• Provide both PDF (vector) and high-resolution PNG outputs

Output: figures/ directory with PDFs, generation scripts, and LaTeX snippets.

Checkpoint:

Figures complete:
- Auto-generated: [list]
- Manual (need your input): [list]
- Rate verification plots: [Y/N]
- Rate comparison table: [Y/N]

[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 /stat-paper-write to generate section-by-section LaTeX:

/stat-paper-write "PAPER_PLAN.md"

What this does:

• Write each section with statistics-appropriate style and structure
• Set up venue-specific template (IMS, JASA, Biometrika, COLT, etc.)
• Write assumption blocks with labels and discussion
• Write theorem environments with proof sketches
• Insert rate comparison table
• Write simulation studies with proper DGP specification
• Write supplement with full proofs
• Build references.bib with verified entries
• Clarity pass with statistics-specific checks
• Claude subagent reviews quality

Output: paper/ directory with main body + paper/supplement/ with proofs.

Checkpoint (theory/methodology):

LaTeX writing complete:
- Main body sections: [N] written
- Supplement sections: [N] written
- Theorems: [N] stated + [N] proof sketches in main body
- Assumptions: [N] labeled (A1)-(AN)
- Rate comparison table: YES
- Citations: [N] unique keys
- Simulation DGPs: [N]

Shall I proceed with compilation?

Checkpoint (application paper):

Application paper LaTeX complete:
- Main body sections: [N] written (§6 Application at [X] pages — centerpiece)
- Supplement sections: [N] written (proofs in supplement)
- §2 Data and Background: [X] pages with [N] EDA figures, [N] tables
- §4 Theory: [1-2] theorems with light discussion
- §5 Simulations: [N] DGPs (informed by §2 data characteristics)
- §6 Application: main analysis + comparison + validation + interpretation
- Comparison methods used: [N] including [N] domain-standard methods
- Data availability statement: present
- Code availability statement: present
- Citations: [N] unique keys (statistical + domain)

Shall I proceed with compilation?

Phase 4: Compilation

Invoke /paper-compile to build the PDF:

/paper-compile "paper/"

What this does:

• latexmk -pdf with multi-pass compilation
• Auto-fix common errors
• Compile both main paper and supplement
• Post-compilation checks: undefined refs, page count

Output: paper/main.pdf + paper/supplement/supplement_main.pdf

Checkpoint:

Compilation complete:
- Main paper: SUCCESS ([X] pages)
- Supplement: SUCCESS ([Y] pages)
- Undefined references: [N]
- Undefined citations: [N]

Shall I proceed with the improvement loop?

Phase 5: Auto Improvement Loop with Codex External Dialogue

The improvement loop combines internal Claude review (every round, fast structural fixes) with external Codex MCP dialogue (final round, senior-statistician depth at GPT-5.5 xhigh).

The dialogue principle applies in Phase 5. Codex's review is one senior reader's opinion, not a directive. The loop discusses with Codex until both sides converge on what the draft needs, not applies Codex's feedback wholesale. Read ../stat-shared-references/stat-codex-dialogue.md before starting Round 2.

Default flow when REVIEW_MODE = both:

Round 1 (Claude internal review):

• Claude subagent reviews the compiled draft.
• Implements CRITICAL and MAJOR fixes.
• Recompiles to main_round1.pdf and supplement_round1.pdf.

Round 2 (Codex external dialogue):

• Initial Codex MCP review at GPT-5.5 with xhigh reasoning (see stat-paper-write Step 6 Pass B for the prompt template).
• For each Codex criticism, decide per ../stat-shared-references/stat-codex-dialogue.md: accept, push back via mcp__codex__codex-reply, or log disagreement.
• Apply accepted criticisms only (not all criticisms).
• Recompile to main_round2.pdf and supplement_round2.pdf.

Optional Round 3 (extended Codex dialogue):

• For high-stakes submissions, continue mcp__codex__codex-reply on the most important remaining issues from Round 2.
• The goal is convergence, not unanimity. Stop when both sides agree, when disagreements are documented, or after diminishing returns (typically after 2 to 4 dialogue rounds total).
• Apply additional accepted criticisms.
• Recompile to main_round3.pdf.

Codex prompt template for the improvement loop:

mcp__codex__codex:
  model: gpt-5.5
  sandbox: read-only
  config: {"model_reasoning_effort": "xhigh"}
  prompt: |
    You are a senior statistician reviewing this manuscript at the
    standard of [VENUE]. Paper type: [theory/methodology/application].

    Below is the current draft after one round of automated improvement.

    [paste main body and supplement]

    Please provide:
    1. Top-line verdict: ready to submit, needs minor revision,
       needs major revision, or fundamental problems?
    2. Three to five highest-priority remaining issues, with the
       minimum fix for each
    3. AI-tell audit: count em-dashes, colons (outside lists/captions),
       semicolons, formulaic openings, empty connectives ("Importantly,",
       "Notably,"), watchwords (delve, pivotal, landscape, etc.)
    4. Main-supplement independence check: any broken cross-references?
       Theorems properly restated in supplement?
    5. Figure design audit: any titles inside figures? Any captions
       not self-contained?
    6. Mock referee report at the venue's standard

    Be direct. Senior statisticians prefer hard feedback in a measured
    voice.

After receiving the review, save the threadId and use mcp__codex__codex-reply to dig into the highest-priority items:

• "Issue 1 is [X]. Please write the specific replacement text for the affected lines."
• "Please write the mock referee response we should expect at [venue]."
• "If we cannot run [requested additional simulation] before submission, what is the next-best mitigation in writing?"

Backward-compatible fallback:

If Codex MCP is unavailable, the loop falls back to /auto-paper-improvement-loop "paper/" with Claude subagent reviews for both rounds.

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

Statistics-specific review focus (passed to reviewer):

For theory/methodology papers:

• Theorem precision: are statements self-contained and all conditions listed?
• Assumption coverage: are all assumptions used and discussed?
• Rate optimality: is the comparison with prior bounds explicit and fair?
• Proof sketch quality: do they provide genuine insight or just defer?
• Simulation rigor: DGP specification, standard errors, rate verification
• Notation consistency across main body and supplement
• Measured tone: no overclaiming, proper hedging where needed

For application papers (additionally or instead):

• Scientific question clarity in the Introduction
• Data and Background section completeness (EDA, descriptive table, statistical challenges)
• Methodology scoping appropriately for the problem (not over-generalized)
• Theory weight in main body (1-2 theorems max; rest in supplement)
• Simulation DGPs informed by real data characteristics
• Application section depth (multiple sub-analyses, multiple pages, dominant section)
• Substantive findings clarity and interpretation
• Comparison with domain-standard methods
• Validation rigor (holdout, CV, sensitivity)
• Practical recommendations in Discussion
• Data and code availability statements present
• Reproducibility addressed concretely
• Dual-audience accessibility (statisticians + domain readers)
• Self-contained figure captions in §6
• Measured tone, no overclaiming of findings

Two rounds of review → fix → recompile.

The reviewer prompt should explicitly invoke /stat-polishing standards. The polishing checks the reviewer applies in each round include:

• Punctuation discipline (em-dashes cut, colons restricted, semicolons reduced)
• AI-template removal (no formulaic openings, no empty connectives, no watchwords, no hedge-stacking, no generic conclusions)
• COPSS-style scholar voice (confidence without hype, plain verbs, connective restraint, mathematical precision over praise)
• Paragraph and bullet discipline (4-8 sentence paragraphs, bullets only where appropriate)
• Figure design rules (no titles, self-contained captions, no legend overlap)
• Main-supplement separation (no broken cross-references, supplement is self-contained)

Output: PDFs for comparison + PAPER_IMPROVEMENT_LOG.md.

Dispatching review findings. When a round surfaces a substantive concern (not a line edit), send it to its owner with ../stat-shared-references/stat-review-routing.md rather tha

…

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: gyf9712
• Source: gyf9712/stat-writing-skills
• License: MIT

Install and usage instructions live in the source repository linked above.