AI Research Writing Skill
δΈζθ―΄ζ
Turn an ML/AI research repo into an evidence-backed, build-ready conference paper draft.
Point your coding agent at code, experiment logs, notes, and a venue template. This skill helps it produce an auditable LaTeX draft and submission package: story, claim-evidence map, verified citations, figures, reviewer-style critique, rebuttal risks, and build notes.
Claim-evidence engineering, not prose generation.
Every major claim should trace to code, results, notes, or verified citations.
End-to-End Demo
Use AI Research Writing Skill to write a complete system paper about this repository itself.
Treat ai-research-writing-skill as the research artifact.
Inspect SKILL.md, references/, scripts/, templates/, examples/, and README.
Create paper_story.md, claim_evidence_map.md, literature positioning, verified citations, ICML-style method figures/tables, and a build-ready ICML LaTeX paper under examples/paper-about-ai-research-writing-skill/paper/.
Do not invent performance numbers. Use repository facts as evidence.
The example already includes the expected final paper package, so you can inspect what an end-to-end output looks like:
- Read the generated ICML-style paper PDF:
paper/main.pdf
evidence/repository_inventory.md: repository facts used as evidence.paper_story.md and claim_evidence_map.md: story and claim boundaries.literature/positioning.md and citation_verification.md: related-project positioning.paper/figures/method_overview.tex and paper/tables/*.tex: ICML-style paper assets.paper/main.tex: complete paper draft about this project.paper/main.pdf: compiled paper for quickly judging output quality.build_check.md: compilation command, expected result, and residual risks.evaluation/auto_research_handoff_results.json: real CLI-to-validator compatibility result between auto-research and this skill.
User Journeys
Use the skill at different stages of a research project. Start with the journey that matches what you already have.
| Journey | You have | Ask the agent to do |
|---|
| Brainstorm and plan | A topic, rough idea, or possible method | Clarify the thesis, research gap, contribution boundary, evidence needed, and next experiment/writing plan. |
| Repo to full paper | Ideas, method design, code, notes, experiment logs, partial results, or a venue template | Build paper_story.md, claim_evidence_map.md, literature positioning, figures, tables, BibTeX, LaTeX draft, and build checks. |
| Draft review | A complete or partial paper draft | Act as a skeptical reviewer: write reviewer-style comments, identify rejection risks, and turn major issues into concrete edits. |
| Targeted revision | A draft plus known weaknesses, reviewer feedback, or a section to improve | Revise the section or whole paper while preserving claim-evidence boundaries and avoiding unsupported stronger claims. |
| Figures and tables | Results, logs, CSVs, method notes, or a rough figure idea | Produce figure/table plans, generated overview or method figures, deterministic result plots/tables, captions, and LaTeX wiring. |
| Submission readiness | A near-final paper package | Run citation, marker, build, venue, reviewer-risk, checklist, and packaging checks before submission or Overleaf/Git handoff. |
Copyable prompts:
I only have a rough idea. Use AI Research Writing Skill to brainstorm the paper story, identify the research gap, define claims to make/avoid, and create a concrete plan for evidence, experiments, figures, and writing.
I have code, notes, method design, and some experiment logs/results. Use AI Research Writing Skill to generate a complete paper package: paper_story.md, claim_evidence_map.md, literature positioning, verified BibTeX, figures, tables, LaTeX draft, and build_check.md.
I have a paper draft. Use AI Research Writing Skill as a skeptical ICML reviewer: write detailed reviewer comments, identify rejection risks, and convert the high-risk issues into concrete revisions.
I have a draft and want to revise it. Use AI Research Writing Skill to improve the paper section by section, preserve supported claims, weaken unsupported claims, fix citations, improve figures/tables, and update the LaTeX package.
Why this skill
| Typical AI paper help | AI Research Writing Skill |
|---|
| Fluent paragraphs from memory | Claims mapped to repo evidence |
| Citations guessed or invented | BibTeX from arXiv / DOI / Semantic Scholar |
| Figure βplansβ that never ship | Generated overview/method figures + deterministic result plots |
| Stops at an outline | Concrete artifacts: paper_story.md, claim_evidence_map.md, references.bib, figure files |
| Generic writing tips | Reviewer-style comments, rejection-risk diagnosis, venue checklists, build & packaging gates |
Venue profiles and audited template sources: NeurIPS, ICML, ICLR, CVPR, ICCV, ECCV, ACL, AAAI, COLM, and related ML/AI conferences. Templates are fetched from recorded official sources rather than redistributed when licensing is unclear. Always verify official author instructions before submission.
Before / After
| Before using this skill | After using this skill |
|---|
| Scattered notes, logs, and half-written sections | paper_story.md and scoped task packets |
| Claims that sound plausible | claim_evidence_map.md with evidence status |
| Candidate citations from memory | BibTeX fetched or verified from authoritative metadata |
| Figure ideas in prose | Generated concept figures and deterministic result plots |
| "Looks done" | Reviewer-style critique, build, TODO, citation, and submission checks |
What you get
End-to-end coverage from idea to camera-ready:
| Stage | Outputs |
|---|
| Story | Thesis, gap, contributions, claims to avoid |
| Evidence | claim_evidence_map.md tied to code / logs / tables |
| Writing | Abstract, Intro, Related Work, Method, Experiments, Limitations, Conclusion |
| Literature | Local corpus, positioning, verified references.bib |
| Figures | Plan + assets: generated overview/method diagrams; deterministic plots for numbers |
| Review | Reviewer-style comments, rejection-risk diagnosis, self-review, concrete revision plan |
| Submit | Venue checklist, LaTeX build check, TODO/citation audit, packaging |
Quality gates (built in)
The skill enforces checkpoints agents must not skip:
- Evidence β numbers from data/logs/scripts, not image models
- Story β no full draft before contributions are explicit
- Literature β positioning before Related Work prose
- Citation β no unverified BibTeX without a visible placeholder
- Figures β concept diagrams via image generation (default); TikZ/SVG only as optional reference
- Reviewer β high-severity objections addressed before βdoneβ
- Build β compile or document why not
Installation
Simple rule: the only canonical entrypoint is the root SKILL.md.
The rest of the repository is supporting material loaded as needed: references/, deterministic scripts/, audited template sources, and an optional end-to-end example.
Agent-Assisted Install
Ask your agent to install this repository as a local skill:
Install https://github.com/jin-s13/ai-research-writing-skill as a local skill.
Use the repository root SKILL.md as the canonical entrypoint.
If your platform needs a skills directory, copy or symlink the whole repository there.
Manual Install
git clone https://github.com/jin-s13/ai-research-writing-skill.git
# Then copy or symlink this repository into your agent's local skills directory.
You can also use it without installing by opening this repository and asking:
Use the AI Research Writing Skill in this repository.
Follow SKILL.md and load only the relevant references for my task.
Repository layout
ai-research-writing-skill/
βββ examples/ # Minimal demo paper repo
βββ SKILL.md # Canonical agent entrypoint
βββ references/ # Workflow, writing, citations, figures, venues, review
βββ scripts/ # Claims, citations, TODOs, build-log, camera-ready checks
βββ templates/ # Audited official-source manifest; no vendored author kits
βββ README.zh-CN.md
Start here when digging in:
Helper scripts
python3 scripts/extract_claims.py main.tex > claim_evidence_map.md
python3 scripts/check_citations.py main.tex references.bib
python3 scripts/verify_citations.py /path/to/paper-project --mailto you@example.org
python3 scripts/check_citation_lock.py /path/to/paper-project
python3 scripts/check_todos.py main.tex checklist.tex references.bib figures
python3 scripts/check_numeric_evidence.py /path/to/paper-project
python3 scripts/check_research_handoff.py /path/to/handoff-project --require-unblocked
python3 scripts/parse_build_log.py main.log
python3 scripts/camera_ready_check.py main.tex
python3 scripts/fetch_template.py icml2026 --output /path/to/paper/.venue-template
python3 scripts/record_build.py /path/to/paper-project --run
python3 scripts/research_quality_gate.py /path/to/paper-project
More: scripts/README.md.
Safety & hygiene
- Conference author kits are fetched from audited official sources and verified by SHA-256 when pinned; confirm current venue rules before submitting.
- Do not commit private PDFs, proprietary logs, API keys, or reviewer-confidential material.
Acknowledgements
This project is inspired by and builds on the excellent research-writing and figure-making projects below. It does not try to replace them; it narrows their ideas into one opinionated workflow: turn an ML/AI research repository into an evidence-backed, build-ready conference paper package.
| Project | What it is excellent at | How this project is different |
|---|
| Master-cai/Research-Paper-Writing-Skills | A compact skill package for ML/CV/NLP paper writing, adapted from research-writing notes into reusable agent skills. | Adds a full repo-to-paper production contract: inventories, claim-evidence maps, verified BibTeX, figure assets, build checks, and submission packaging. |
| Norman-bury/research-writing-skill | A broad, multi-platform research-writing assistant for thesis writing, chapter workflows, literature review, LaTeX output, and process tracking. | Specializes in AI conference papers from code, logs, experiments, and venue templates instead of general thesis/chapter writing. |
| Orchestra-Research/AI-research-SKILLs | A comprehensive AI research and engineering skills library for agents, spanning the broader research lifecycle from idea to paper. | Focuses on one deep vertical: paper-writing execution and submission readiness for an existing ML/AI research repo. |
| Yuan1z0825/nature-skills | Nature/CNS-style academic writing, polishing, reviewer response, data availability, citation, and publication-quality figure workflows. | Targets ML/AI conference workflows and templates such as NeurIPS, ICML, ICLR, CVPR, ICCV, ECCV, ACL, AAAI, and COLM. |
| ChenLiu-1996/figures4papers | High-quality Python scripts and examples for publication figures in top AI conferences and journals. | Integrates figure planning into a larger paper pipeline: figures are tied to claims, evidence, captions, LaTeX references, and submission checks. |
In short: the related projects provide writing wisdom, broad skill ecosystems, Nature-style publication craft, or figure-making expertise. This project packages those inspirations into a claim-evidence-engineering workflow for AI paper agents working inside real research repositories.
License
MIT License.
