Arxiv Paper Writer Agent Skill

ML/AI Review Paper Workflow (IEEEtran template)

When to Use

• ML/AI review papers for arXiv (main text ~6-10 pages; references excluded)
• LaTeX + BibTeX workflow with verified citations
• Citation validation/repair on existing LaTeX projects

When NOT to Use

• Novel experimental research papers (this is a review workflow)
• Non-academic documents

Inputs

• Topic description (required)
• Constraints: venue, page limit, author/affiliations (optional)
• Existing project path for citation validation (optional)

Outputs

• main.tex (LaTeX source)
• ref.bib (verified BibTeX entries)
• IEEEtran.cls
• plan/<timestamp>-<slug>.md, issues/<timestamp>-<slug>.csv
• Figures/tables; main.pdf
• notes/research-snapshot.md (Gate 0 research snapshot)
• notes/literature-notes.md (optional per-citation notes)
• notes/arxiv-registry.sqlite3 (arXiv metadata/BibTeX cache)

Conventions: run python3 scripts/... from this skill folder (where scripts/ lives); <paper_dir> is the paper/project root (contains main.tex, ref.bib, plan/, issues/, notes/). Paths like plan/... are under <paper_dir>. Papers are created under --out <workspace_dir> (your working directory, never inside this skill folder); pass the same --out (and --name, if used) to both bootstrap stages. For arXiv discovery/metadata/BibTeX, use scripts/arxiv_registry.py (no ad-hoc curl/wget).

Gated Workflow

> Tip: Run python3 scripts/<script>.py --help before use.

> Open reference files only when a step calls them out.

Non-Negotiable Rules

• No prose in main.tex until plan approved AND issues CSV exists.
• First deliverable: research snapshot + outline + clarification questions + draft plan.
• Use plan + issues tracking for all new papers; do not opt out.
• Issues CSV is the execution contract; update Status and Verified_Citations per issue, and add/split/insert issue rows when scope grows (do not do untracked work).
• Template is fixed: use IEEEtran two-column layout (assets/template/IEEEtran.cls).

Treat two-column width as a layout constraint (use two-column floats when needed).

Gate 0: Research Snapshot + Draft Plan

• Confirm constraints (venue, page limit, author block, date range).
• Translate the topic into search keywords and run a light discovery pass:

10-20 key papers (see references/research-workflow.md). After step 4 (once <paper_dir> exists), cache arXiv discovery with arxiv_registry.py search and save the snapshot to notes/research-snapshot.md.

• Propose 2-4 candidate titles aligned to the topic.
• Scaffold the project folder and draft plan:

   python3 scripts/bootstrap_ieee_review_paper.py --stage kickoff --topic "<topic>" --out <workspace_dir>

This copies LaTeX templates from assets/template/; plan/issues are generated from templates in assets/.

Initialize arXiv registry (once): python3 scripts/arxiv_registry.py --project-dir <paper_dir> init.

• Create a framework skeleton in main.tex

(section headings + 2-4 bullets per section + seed citations; no prose).

• Update the plan file to reflect the framework, proposed titles, and section/subsection plan.
• Compile early: python3 scripts/compile_paper.py --project-dir <paper_dir>

Fix any Overfull \hbox warnings (see Layout Hygiene below).

• Return to user:
• Proposed outline (5-8 sections, 2-4 bullets each)
• Planned visualizations (5+) mapped to sections (see references/visual-templates.md)
• Clarification questions
• STOP until user approves.

Gate 1: Create Issues CSV (after approval)

• Check kickoff gate in plan: - [x] User confirmed scope + outline in chat.
• Create issues CSV (script refuses if gate unchecked):

   python3 scripts/bootstrap_ieee_review_paper.py --stage issues --topic "<topic>" --out <workspace_dir> --with-literature-notes

• Validate:

   python3 scripts/validate_paper_issues.py <paper_dir>/issues/<timestamp>-<slug>.csv

• If literature notes are enabled, keep short summaries and (optional) abstract snippets to avoid re-search.
• The plan may evolve; add/split/insert issues as needed, re‑validate after edits, and keep going until all issues (including inserted ones) are DONE or SKIP (when feasible, in the same run).

Phase 2: Per-Issue Writing Loop

For each writing issue in the CSV:

• If an issue balloons (new figure, new subsection, new benchmark set, or a large QA fix), split/insert new issue row(s) (e.g., W6a, Q5) before proceeding; re-run python3 scripts/validate_paper_issues.py <issues.csv>; keep going until all issues are DONE/SKIP.
• Research: 8-12 section-specific papers.
• Write: Never 3 sentences without citations; varied paragraph rhythm

(see references/writing-style.md).

For section intent and structure, use references/template-usage.md.

• Visualize: Match content triggers (see references/visual-templates.md).

Prioritize single-column sizing; use double-column spans only when necessary (see Layout Hygiene).

Cite externally sourced figure content.

• Verify: Web search + open source page (and PDF if available) before adding to ref.bib.

For arXiv entries, append BibTeX via python3 scripts/arxiv_registry.py --project-dir <paper_dir> export-bibtex <arxiv_id> --out-bib <paper_dir>/ref.bib.

• Update: Mark issue DONE with Verified_Citations count.
• Compile after meaningful changes; fix Overfull \hbox before marking DONE.

Phase 2.5: Rhythm Refinement

After all writing issues are DONE, refine prose section-by-section using the latex-rhythm-refiner skill. This step varies sentence/paragraph lengths and removes filler phrases while preserving all citations.

Phase 3: QA Gate

• Run internal QA checklist (see references/quality-report.md).
• Compile; ensure no Overfull \hbox warnings in main.log.
• Deliver main.tex, ref.bib, figures, and main.pdf.

Existing Paper Workflow (No Re-Scaffold)

If a paper folder already exists, do NOT rerun scaffold:

# Create plan
python3 scripts/create_paper_plan.py --topic "<topic>" --stage plan --output-dir <paper_dir>
# STOP for approval, then check kickoff gate box
# Create issues (use timestamp/slug from plan filename/frontmatter)
python3 scripts/create_paper_plan.py --topic "<topic>" --stage issues --timestamp "<TS>" --slug "<slug>" --output-dir <paper_dir> --with-literature-notes

Citation-Validation Variant

• Treat provided path as LaTeX project root.
• Follow references/citation-workflow.md.
• Use references/bibtex-guide.md for BibTeX rules if entries need repair.
• Deliver validation report and corrected ref.bib if requested.

Success Criteria

Compilation: python3 scripts/compile_paper.py --project-dir <paper_dir> (exit 0, no "Citation undefined" warnings). Use --report-page-counts for main-text page count.

Quality Metrics:

• 6-10 pages of main text (references excluded)
• 60-80 total citations (8+ per section)
• 100% citation verification rate
• 70%+ citations from last 3 years
• 5+ visualization types
• All issues DONE or SKIP

Safety & Guardrails

• Never fabricate citations or results; add TODO and ask user if evidence missing.
• Verify every citation via web search + source page (and PDF if available) before adding to ref.bib.
• Confirm before large literature searches.
• Do not overwrite user files without confirmation.
• Issues CSV is the contract; mark DONE only when criteria met.
• No submission bundles unless user requests.

Layout Hygiene

Fix Overfull \hbox warnings before marking issues DONE:

• Figures: start with figure + \columnwidth; switch to figure* + \textwidth if needed
• Tables: prefer p{...} column widths / \tabcolsep over \resizebox
• Equations: use split, multline, aligned, or IEEEeqnarray for line-breaking

Issues CSV Schema

| Phase | Issues |

|-------|--------|

| Research | Rx: discovery, scaffolding, framework, viz planning |

| Writing | Wx: each section with target citations and visualization |

| Refinement | RFx: apply latex-rhythm-refiner skill (after all Wx DONE) |

| QA | Qx: citation verification, QA checklist, compilation, final review |

Status: TODO → DOING → DONE, or SKIP with a reason in Notes. Schema validated by validate_paper_issues.py.

Verified_Citations = the number of unique verified cite keys in the issue's section when it is marked DONE (0 for non-writing issues).