How to use the Rebuttal Pipeline agent skill

What is an agent skill, and what does the rebuttal skill do?

An agent skill is a folder containing a SKILL.md file that gives an AI coding agent such as Claude Code a specific workflow to follow. The rebuttal skill is named rebuttal and is described as Workflow 4 in the source: a submission rebuttal pipeline that runs after external reviews arrive.

The skill parses the reviews, enforces coverage and grounding, drafts a safe text-only rebuttal under venue limits, and manages follow-up rounds. It works in phases: it atomizes each reviewer concern into an issue board with IDs such as R1-C2, classifies each concern by type and severity, and builds a strategy plan before any drafting happens.

• Splits reviews into individual concerns, each tagged with reviewer, round, and severity.
• Checks that every claim traces to the paper, the reviews, or facts you confirmed.
• Checks that every promise is one you have approved.
• Ensures every reviewer concern ends as answered, deferred intentionally, or needing your input.
• Can run an internal stress-test in which a separate AI reviewer critiques the draft.

Install the rebuttal skill

You can install the skill in one of two ways. The first is to download the ZIP file from the skill's page on this directory and unzip it into your personal skills folder at ~/.claude/skills/. To scope it to a single project instead, unzip it into the .claude/skills/ folder inside that project.

The second way is to upload the ZIP in the Claude apps. Open Settings, go to Capabilities, and upload the skill under Skills.

The skill's only file listed in the source is SKILL.md, so the unzipped folder should contain that file.

• Personal install: ~/.claude/skills/
• Project install: .claude/skills/ inside your project folder
• Claude apps: Settings → Capabilities → Skills, then upload the ZIP

Gather what the skill needs before you start

The skill needs four inputs. The paper source can be a PDF, a LaTeX directory, or a narrative summary. The raw reviews can be pasted text, markdown, or a PDF with reviewer IDs. The skill also needs explicit venue rules: the venue name, the character or word limit, whether revised PDFs are allowed, and whether the venue shows one shared response or separate reviewer threads.

If venue rules, the limit, or the rendering mode are missing, the skill is designed to stop and ask before drafting. The skill also needs a reviewer backend for its stress-test step. The default is Codex MCP; the source also lists Oracle and a manual review MCP as alternatives.

• Paper source: PDF, LaTeX directory, or summary
• Raw reviews as text, markdown, or PDF
• Venue name, character or word limit, and whether revised PDFs are allowed
• Reviewer backend for the stress test (default Codex MCP)

Trigger the skill with a natural request

The skill activates when you use phrases such as "rebuttal", "reply to reviewers", "ICML rebuttal", or "OpenReview response", or when you want to answer external reviews safely. You can also pass an argument naming the paper path or review bundle.

The source shows an override form that sets the venue and limit inline. Example requests in plain language:

• "Help me write a rebuttal for these ICML reviews. The limit is 5000 characters."
• "Reply to reviewers for my OpenReview submission. Paper is in the paper/ folder and the reviews are in reviews.md."
• "I need to answer external reviews safely. Make a checklist of every reviewer concern first."

Work through the output files

The skill writes its work into a rebuttal/ folder. The source names several outputs: REVIEWS_RAW.md holds the reviews verbatim, ISSUE_BOARD.md lists each concern, STRATEGY_PLAN.md records the response plan, REBUTTAL_STATE.md tracks the current phase so a session can resume, and PASTE_READY.txt is the plain-text rebuttal. A detailed draft, REBUTTAL_DRAFT_rich.md, can be rendered to HTML.

Review these files before you paste anything. The skill produces text only, so you copy the final response into the venue's system yourself.

• rebuttal/REVIEWS_RAW.md: reviews stored verbatim
• rebuttal/ISSUE_BOARD.md: one entry per atomic concern
• rebuttal/STRATEGY_PLAN.md: how each concern will be answered
• rebuttal/PASTE_READY.txt: the character-counted plain-text rebuttal
• rebuttal/REBUTTAL_DRAFT_rich.md: the detailed draft, rendered to HTML by default

Use quick mode to understand reviews first

If you only want to see what reviewers are asking before deciding how to respond, use quick mode. The source describes QUICK_MODE as running only the parse, atomize, and strategy phases. It outputs ISSUE_BOARD.md and STRATEGY_PLAN.md and stops before drafting or stress-testing.

Quick mode is a sensible first pass when reviews are long or when several co-authors need to agree on the approach.

Safety notes: scripts, network, and what the skill will not do

The skill runs scripts and uses the network. Its allowed tools include shell commands and calls to reviewer MCP backends, so it sends prompts to whichever reviewer backend you select. Check which backend is configured before you start, and do not send confidential material to a backend you have not approved.

The skill does not run new experiments, generate new theorem claims, edit or upload a revised PDF, or submit to OpenReview, CMT, or HotCRP. If a reviewer asks for new evidence, the skill flags an evidence gap for you to handle. Any new result or commitment must come from you and be confirmed before the skill uses it.

• Runs scripts: yes. Uses network: yes.
• Stress-test prompts go to the reviewer backend you choose.
• Does not submit anything; you paste and upload the text yourself.
• Any promise in the rebuttal must be one you have approved.

Tips for a stronger rebuttal

Give the skill complete reviews and an explicit character limit, since missing limits stop the workflow. Confirm any new result yourself and tell the skill about it, so it can be cited as user-confirmed evidence rather than guessed. Read the issue board before drafting to make sure no concern is missing.

Use the stress test as a critique, not a verdict. Fix any issue it raises that is substantive, and keep your own judgment about what to concede.

• Paste the full reviews, including reviewer IDs.
• State the venue's limit and whether revised PDFs are allowed.
• Check ISSUE_BOARD.md for missing concerns before drafting.
• Only approve promises you can actually keep.