📝 rebuttal-skills

Draft a grounded rebuttal to your paper's reviews — with the extra experiments actually run in your workspace

_{Grounded in 205,988 real peer reviews · ICLR · ICML · NeurIPS · COLM · 2023–2026}

You got your reviews back. Now you have three days, a reviewer demanding a significance test, and a
repo full of code that could answer them.

rebuttaldraft reads the paper and the reviews, finds how authors of 200k+ real papers answered
the same criticism, works out which experiments would actually settle each concern, runs them in your
workspace, and writes a paste-ready reply per reviewer, filled with the numbers it just measured.

✨ Why this is different

🔬 Runs the experiments. Not "we will report this in the revision" — an actual table, from an
actual run, in your actual repo. This is the whole point, and it is what a chat window cannot do.
📚 Grounded in real references. Every concern is matched against 200k+ reviews and the rebuttals
that answered them, so the draft mirrors what actually worked, not what sounds good.
🔒 Your paper never leaves your machine. Only short, abstract search queries reach the API. The
paper text, the review text, and your results stay local.
🚫 Never invents a number. Every figure in the draft traces to a real artifact or to your
paper. An experiment that could not run is answered honestly, not with a plausible-looking result.

🔁 How it works

Stage	What happens
0 Intake	Finds your paper (PDF / TeX / md) and the reviews
1 Concerns	Extracts every distinct concern per reviewer, with severity
2 Grounding	Searches the corpus, keeps the one best precedent per concern
3 Experiments	Inventories what you already have, designs what is missing, and runs it
4 Draft	Writes the per-reviewer reply, tables and all, from the real results
5 Critic	Re-reads the draft as a skeptic, checks every number, revises

📦 Installation

Claude Code

/plugin marketplace add yjoonjang/rebuttal-skills
/plugin install rebuttal

One install, two skills: rebuttal:rebuttaldraft (drafting) and rebuttal:reviewsearch (search).

Codex

Clone and symlink into Codex's skills directory, then restart Codex:

git clone https://github.com/yjoonjang/rebuttal-skills.git ~/.codex/rebuttal-skills
mkdir -p ~/.agents/skills
ln -s ~/.codex/rebuttal-skills/skills ~/.agents/skills/rebuttal

Both skills are then discovered natively, in every workspace. Details: .codex/INSTALL.md.

🚀 Use

Put your paper and your reviews in your project, then just ask:

"help me write a rebuttal to these reviews"

"respond to reviewer 2"

It will confirm the files, extract the concerns, show you the experiment plan, and wait for your
approval before running anything. Output lands in rebuttal/:

rebuttal/
├── R1.md, R2.md, R3.md         # per-reviewer, paste-ready: Overview → Draft → Guide
└── experiments/
    ├── RESULTS.md              # every number, with its provenance
    ├── grounding.md            # the precedent cases the drafts learned from
    └── ...                     # run logs and artifacts

🔍 ReviewSearch

The drafts are grounded in ReviewSearch, hybrid semantic search over 205,988 peer reviews and
their author rebuttals. It ships as its own skill, and it is useful on its own.

ReviewSearch demo

Ask naturally:

"find reviews asking for a significance test"

"how did papers rebut novelty concerns?"

Or run the script directly (Python 3, no dependencies):

python skills/reviewsearch/scripts/search_reviews.py "missing ablation on learning rate" --top-k 5

Flag	Description
`--top-k N`	Number of results to return (default `10`)
`--accepted-only`	Restrict to accepted papers
`--year-min YYYY`	Inclusive earliest year (2023–2026)
`--year-max YYYY`	Inclusive latest year (2023–2026)

Each result carries venue, year, title, decision, summary, concern (the reviewer's
weaknesses and questions), rebuttal (the authors' response), and score.

💡 Try it in the browser: live demo.
Override the backend with RS_API_URL. The free Space cold-starts (~1 min) after being idle — retry once.

📚 Attribution

Data: peer reviews from OpenReview (ICLR / ICML / NeurIPS / COLM), licensed CC-BY-4.0.

📄 License

Plugin code: MIT — see LICENSE.