ai-research-writing-skill

AI Research Writing Skill

中文说明

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.

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

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

What you get

End-to-end coverage from idea to camera-ready:

Quality gates (built in)

The skill enforces checkpoints agents must not skip:

• Evidence — numbers from data/logs/scripts, not image models
• Story — no full draft before contributions are explicit
• Literature — positioning before Related Work prose
• Citation — no unverified BibTeX without a visible placeholder
• Figures — concept diagrams via image generation (default); TikZ/SVG only as optional reference
• Reviewer — high-severity objections addressed before “done”
• Build — 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/
├── examples/             # Minimal demo paper repo
├── SKILL.md              # Canonical agent entrypoint
├── references/           # Workflow, writing, citations, figures, venues, review
├── scripts/              # Claims, citations, TODOs, build-log, camera-ready checks
├── templates/            # Audited official-source manifest; no vendored author kits
└── 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.

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.