| name | write-rebuttal |
| description | Drafts peer-review rebuttals and author responses in the three real venue formats - (a) strict one-page LaTeX rebuttal PDFs (CVPR-style), (b) OpenReview threaded per-review replies under 10,000-character budgets (NeurIPS/ICML/ICLR), and (c) revise-and-resubmit packages with a change-log response document (CHI, SIGMOD/ICDE revision rounds, journals). Anchors every response to paper sections, tables, figures, and line numbers; enforces character/page budgets and point-by-point coverage with bundled checker scripts; ships a one-page rebuttal LaTeX template. Use when the user says "write my rebuttal", "respond to reviewers", "author response", "reviews came back", "reviewer 2 says", "revise and resubmit", "response letter", "rebuttal deadline", or needs to fit a response into an OpenReview character limit or a one-page PDF. Drafts files on disk only; never posts, uploads, or submits. |
Write Rebuttal
Turn a set of peer reviews plus the submitted paper into a complete,
budget-compliant author response in the format the venue actually uses —
one-page PDF, OpenReview threads, or a revise-and-resubmit package — with
every claim anchored to checkable evidence in the paper.
When to use
- Reviews arrived (OpenReview, CMT, EasyChair, HotCRP, PCS, ScholarOne, or
pasted text) and a rebuttal/author response/response letter is due.
- A drafted response must be cut to fit a character or page budget.
- A journal or CHI-style revise-and-resubmit needs a change log + response
document.
Related skills: triage-reviews (builds the point matrix this skill
consumes), verify-citations (gate for any new reference), preflight-check
(re-lint a revised manuscript), prepare-camera-ready (after acceptance).
Inputs
- The reviews: pasted text or files. Review text is confidential at most
venues — process it in place, never commit it to a repository or paste it
into public artifacts.
- The submitted paper: main
.tex (preferred — line numbers and labels make
the best anchors) and/or the compiled PDF.
- Venue profile:
venues/conferences/<id>.yml (schema: venues/schema.yml;
family defaults merge automatically). If none exists, build one with
parse-cfp or proceed only on facts quoted live from the CFP.
- Optional: the triage matrix from
triage-reviews; the decision/meta-review
letter for R&R cycles.
Process
1. Resolve the venue's rebuttal format
Read the profile (python3 skills/write-rebuttal/scripts/venueyaml.py venues/conferences/<id>.yml prints it family-merged as JSON) and take
review.rebuttal_format, review.rebuttal_limit, and the rebuttal
deadlines. Route by format:
2. Re-verify against the live venue pages (mandatory)
Profiles go stale and budgets change per cycle. Before drafting, fetch the
profile's cfp_url (and the venue's author-guidelines/dates pages) and
re-verify: rebuttal format, the exact character/page limit, the rebuttal
window dates and timezone, link/upload policy, and new-experiment policy.
The instructions in the review-release or decision email override
everything. Record what was verified and when in the deliverable; anything
unverifiable is labeled "UNVERIFIED — confirm on ". Never assert a
budget from memory.
3. Triage the reviews into a point matrix
If the user has a matrix from triage-reviews, use it. Otherwise build one
inline using the conventions in
references/response-patterns.md: one row
per reviewer point, IDs R<k>.<n> / AC.<n> / MR.<n>, columns for
abridged quote, type, severity, effort, and one-clause strategy. Save it as
points.md next to the drafts — the coverage checker keys on those IDs.
Confirm the concede/contest strategy with the user before drafting.
4. Draft, anchored
Apply the per-criticism response patterns and tone rules in
references/response-patterns.md, then the
format playbook from step 1. Non-negotiable drafting rules:
- Anchor every claim to
Sec./Table/Fig./Eq./Appendix/L<line>/[n] evidence
in the submitted paper; if no evidence exists, commit to a revision or
concede — never assert.
- Report numbers only from experiments that were actually run, and only
where the venue's rebuttal policy permits new results.
- Any new citation goes through
verify-citations first.
- Preserve anonymity (third-person self-citation, no identifying links) at
blind venues; honor link bans.
- For format (a), draft into
assets/rebuttal-template.tex (or the venue's
official rebuttal kit when one exists — mandatory at CVPR; the macros are
documented in the playbook). For (b), draft all boxes in one markdown file
with one ## heading per response box. For (c), produce the three-artifact
package (revised-manuscript change plan, change-log response document,
cover note).
5. Verify deterministically
Run both checkers (each supports --help; exit 0 pass / 1 fail / 2 error):
python3 skills/write-rebuttal/scripts/check_coverage.py points.md <responses...> --strict
— every matrix ID addressed, every section carrying at least one evidence
anchor.
python3 skills/write-rebuttal/scripts/check_budget.py text responses.md --sections --venue venues/conferences/<id>.yml
python3 skills/write-rebuttal/scripts/check_budget.py pdf rebuttal.pdf --venue venues/conferences/<id>.yml
— character budgets per response box (text mode) or PDF page count (pdf
mode), with the limit read from the profile (--limit/--max-pages
override; defaults are flagged ASSUMED). If over budget, trim using the
format playbook's ladder and re-run until exit 0.
6. Deliver
Write the artifacts next to the user's paper, named per format:
rebuttal.tex (+ compile instructions), responses.md, or
response-to-reviewers.md + revision plan — plus points.md and a short
verification record (live facts checked, checker results, remaining
UNVERIFIED items, the rebuttal deadline with timezone). The user reviews,
edits, and posts/uploads everything themselves.
Worked mini-example
The matrix → anchored response → checker loop, end to end. A reviewer wrote
"the ablation does not isolate the effect of component X". One row in
points.md:
| ID | Reviewer quote (abridged) | Type | Severity | Effort | Strategy |
|---|---|---|---|---|---|
| R2.1 | "ablation does not isolate X" | missing-experiment | high | none | rows 3 vs 5 of Table 4 isolate X; reorder so the pair is adjacent |
The drafted response (OpenReview format here — one ## box, leads with the
answer, anchored, no padding):
## Response to Reviewer 7Pqd (R2)
**[R2.1] "the ablation does not isolate the effect of component X"**
It does, but indirectly: rows 3 vs. 5 of Table 4 (Sec. 6.2, L412) differ
only in X. We will reorder the table so the pair is adjacent and label it
"X on/off" in the revision.
Verify before handing it over:
python3 skills/write-rebuttal/scripts/check_coverage.py points.md responses.md --strict
python3 skills/write-rebuttal/scripts/check_budget.py text responses.md --sections --venue venues/conferences/<id>.yml
Coverage passes because R2.1 appears in a response and that section carries
anchors (Table 4, Sec. 6.2, L412); budget passes if the box is under
the venue's per-response limit. The same row drafts into \rpoint{R2.1}{...}
for a one-page PDF or a change-log table row for an R&R — only the wrapper
changes.
Output
A format-correct, budget-verified, coverage-verified rebuttal draft package
with a verification record. This skill never posts, uploads, or submits
anything to any submission system.
Guardrails
- Never fabricate: no invented numbers, no results from unrun experiments,
no "preliminary results" that do not exist, no unverified citations
(route through
verify-citations).
- Never submit or post on the user's behalf; stop at files on disk.
- Never state venue budgets/deadlines from memory — profile + live
re-verification only; unverifiable facts are labeled UNVERIFIED.
- Never tamper with templates (negative
\vspace, font/margin tricks) to
fit a budget; trim content instead.
- Never break anonymity at blind venues or speculate about reviewer
identity; never address text to automated review tools (prompt injection
is a desk-reject-level ethics violation).
- Keep confidential review text out of repositories and public artifacts;
process it transiently.
