| name | paper-rebuttal |
| description | Guide authors through the full rebuttal process for ML conference/journal reviews. Four phases (Understand → Task → Rebuttal → Revise) producing artifacts under 1-review/. Use when the user has received peer reviews and needs to write rebuttal responses, plan experiments to address reviewer concerns, or prepare a revised manuscript. Trigger: rebuttal, reviewer response, author response, review feedback, camera-ready revision. |
| allowed-tools | Bash, Read, Write, Edit, Glob, Skill |
| metadata | {"version":"0.1.1","last_updated":"2026-07-14","summary":"Guide authors through the full rebuttal process for ML conference/journal reviews. LAW 1: the rebuttal session is a CONSUMER -- it DISPATCHES each point to Agent(haipipe-task-orchestrator-agent) / Agent(haipipe-discovery-orchestrator-agent) as ONE general-language question (the executor scaffolds + names the task-folder), then READS the returned QA file's `- state:` line (answered → quote · working → in progress, do not re-dispatch · superseded-by → follow the chain) and NEVER writes in the bank. History: ./CHANGELOG.md."} |
Skill: paper-rebuttal
Guide authors through the full rebuttal process for ML conference and journal paper reviews.
Four phases from reading reviews to revising the paper, designed for multi-author collaboration.
Use this skill when the user has received peer reviews and needs to write rebuttal responses, plan experiments to address reviewer concerns, or prepare a revised manuscript.
Also use when the user mentions rebuttal, reviewer response, author response, review feedback, or camera-ready revision.
Four-Phase Pipeline
Phase A: Understand <- Read reviews, annotate every concern
Phase B: Task <- Map concerns to rebuttal points → tasks
Phase C: Rebuttal <- Plan strategy + write the rebuttal responses
Phase D: Revise <- Generate revision checklist → hand off to /paper-revise
Each phase produces artifacts in the review directory:
1-review/
+-- A-review-content/ <- Annotated reviews (one per reviewer)
+-- B-rebuttal-task/ <- Point → task mapping, experiment plan
+-- C-rebuttal-writing/ <- Strategy + rebuttal responses + supplementary
+-- D-paper-revision/ <- Revision checklist (→ /paper-revise)
Commands
/paper-rebuttal understand [review-file] -> read + annotate one reviewer
/paper-rebuttal task [review-dir] -> map points → tasks
/paper-rebuttal rebuttal [reviewer-id] -> plan + write one rebuttal response
/paper-rebuttal revise [review-dir] -> revision checklist → /paper-revise
/paper-rebuttal audit [review-dir] -> coverage + chars + consistency
/paper-rebuttal supplementary [dir] -> anonymous GitHub repo + README
Default (no arg): show this command list.
Typical workflow:
1. understand (repeat for each reviewer — this is the foundation)
2. task (after all reviews are annotated)
3. rebuttal (repeat for each reviewer)
4. audit (check all responses together)
5. revise (generate checklist, hand off to /paper-revise)
Annotation Convention
Reviews are annotated inline by coauthors and Claude Code (CC).
Each annotation uses blockquote format with initials:
> {INITIALS}: {comment}
Examples:
> {AU1}: We should admit this limitation honestly.
> {AU2}: Agree. Also cite the ADA 2024 guidelines here.
> CC: Task {task_id} addresses this. Results: {key findings}.
Rules:
- Each coauthor uses their own initials (2-3 letters, free-form)
- CC is reserved for Claude Code responses
- Annotations go directly below the reviewer text they respond to
- Multiple coauthors can annotate the same paragraph
- Language is free-form (English, Chinese, mixed — whatever is natural)
Execution Protocol
Step 0: Parse the subcommand from args.
Step 1: Read ref files FIRST (mandatory before proceeding).
understand: ref/concepts.md
task: ref/concepts.md + ref/formats.md
rebuttal/audit: ref/principles.md + ref/formats.md
revise: ref/principles.md
Confirm: "Loaded: [ref files].
Executing: [subcommand]."
Step 2: Read and follow the function file exactly.
understand -> fn/fn-understand.md
task -> fn/fn-task.md
rebuttal -> fn/fn-rebuttal.md
revise -> fn/fn-revise.md
audit -> fn/fn-audit.md
supplementary -> fn/fn-supplementary.md
Review Directory Structure
The skill produces artifacts in a 4-folder structure inside the paper's review directory (e.g., paper/1-review/):
1-review/
+-- README.md <- Overview + status
+-- HANDOFF.md <- Continuation prompt for new sessions
+-- A-review-content/ <- Annotated reviews (one per reviewer)
| +-- README.md <- Scores + master mapping table
| +-- review-raw.md <- Original reviews (raw from venue)
| +-- review-{reviewer_id}.md <- Review + annotations + paper source refs
+-- B-rebuttal-task/ <- Point → task mapping
| +-- README.md <- Point → task table + status
| +-- probe-plan.md <- Execution order, feasibility, deps
+-- C-rebuttal-writing/ <- Strategy + final responses
| +-- 0-rebuttal-strategy.md <- Reviewer goals, char budgets, table allocation
| +-- rebuttal-{reviewer_id}.md <- Annotated version (with > comments)
| +-- rebuttal-{reviewer_id}-clean.md <- Clean version (ready to paste)
| +-- supplementary/ <- Anonymous supplementary tables
+-- D-paper-revision/ <- LaTeX changes
| +-- revision-checklist.md <- All changes with file:line references
+-- _archive/ <- Old files from before reorganization
File Map
SKILL.md <- you are here (router)
ref/concepts.md <- 5 key concepts for understanding reviews
ref/principles.md <- 10 rebuttal writing principles
ref/formats.md <- venue formats (ICML/NeurIPS/ICLR/journal)
fn/fn-understand.md <- Phase A: read + annotate one reviewer
fn/fn-task.md <- Phase B: points → tasks
fn/fn-rebuttal.md <- Phase C: plan strategy + write one response
fn/fn-revise.md <- Phase D: revision checklist → /paper-revise
fn/fn-audit.md <- coverage + chars + consistency check
fn/fn-supplementary.md <- anonymous GitHub repo setup