rebuttal-skills

Two composable Claude Code skills for writing conference/journal author-response rebuttals (NeurIPS, ICLR, ICML, CVPR/ICCV/ECCV, AAAI, ACL/ARR/EMNLP, IEEE venues, …).

Skill	What it does	When it triggers
write-rebuttal	The overall process: two-audience mindset (reviewers and the Area Chair), itemize → brain-dump → draft → revise, 18 drafting tactics, the neutral-third-party test, rebuttal-experiment planning, plus an advisory lint script.	"Write my rebuttal", "respond to reviewers", "author response", planning rebuttal experiments.
nlp-rebuttal	The per-concern playbook: 28 recurring reviewer-concern scenarios (novelty, "simple combination", weak baselines, marginal gains, missing ablations, no significance, data leakage, no human eval, reproducibility, …), each with a bad-answer anti-pattern and a recommended answer template, routed by a concern table.	"The reviewer says…", "R2 complains…", drafting a reply to one specific review comment.

They compose: write-rebuttal shapes the whole response (structure, ordering, tone, the note to the AC); nlp-rebuttal picks the strategy and template for each individual answer.

Install

Option A — as a Claude Code plugin

/plugin marketplace add yuangao-tum/rebuttal-skills
/plugin install rebuttal-skills@rebuttal-skills

Option B — copy the skills directly

git clone https://github.com/yuangao-tum/rebuttal-skills.git
cp -r rebuttal-skills/skills/write-rebuttal rebuttal-skills/skills/nlp-rebuttal ~/.claude/skills/

(Or copy into .claude/skills/ inside a project to scope them to that project.)

Usage

Once installed, the skills trigger automatically when you work on a rebuttal, or invoke them explicitly:

• "Here are my ICLR reviews — help me write the rebuttal" → write-rebuttal drives the process end-to-end.
• "R2 says our method is just a combination of existing techniques" → nlp-rebuttal routes to the matching scenario (Tip 3) and adapts its template.
• Before posting, run the mechanical check:

python skills/write-rebuttal/scripts/lint_rebuttal.py rebuttal.md

It flags promise-language with no numbers ("we will add…"), unfilled placeholders, missing AC-facing summary, and pre-written outcomes — advisory only.

What's inside

skills/
├── write-rebuttal/
│   ├── SKILL.md                          # process, 18 tactics summary, anti-patterns
│   ├── references/eighteen-tips.md       # all 18 tactics with examples + concern→tip map
│   ├── references/rebuttal-skeleton.md   # structure, micro-templates, AC dashboard tables, checklist
│   ├── assets/rebuttal-template.tex      # compilable LaTeX skeleton (colour-coded reviewers,
│   │                                     #   quote boxes, AC standing + coverage tables)
│   └── scripts/lint_rebuttal.py          # advisory rebuttal linter (stdlib only)
└── nlp-rebuttal/
    ├── SKILL.md                          # concern router (28 scenarios), cross-cutting rules
    ├── references/innovation-theory.md   # tips 1–7: novelty, motivation, theory, limitations
    ├── references/communication-writing.md # tips 8–12: related work, clarity, misunderstandings
    └── references/experiments-evidence.md  # tips 13–28: baselines, ablations, significance, …

Credits

• write-rebuttal is distilled from Devi Parikh, Dhruv Batra & Stefan Lee, How we write rebuttals. All credit for the method goes to the original authors.
• nlp-rebuttal is translated and adapted from MLNLP-World/Paper-Rebuttal-Tips (Chinese original). All credit for the scenario collection goes to the MLNLP community contributors. The upstream repository carries no explicit license; the adaptation here is shared with attribution for research/educational purposes and will be adjusted or removed at the upstream authors' request.

License

This repository is released under the MIT License.