agentleFS
yuangao-tumyuangao-tum/rebuttal-skills/skills/nlp-rebuttal/SKILL.md

nlp-rebuttal

Skill5 starsChanged 4 months agoMITView on GitHub

Scenario playbook for answering a SPECIFIC reviewer concern in an NLP/ML/AI rebuttal — 28 concern types (novelty, simple combination, unclear motivation, weak baselines, marginal gains, missing ablations, no significance, data leakage, no human eval, reproducibility, and more), each with a bad-answer anti-pattern and a recommended-answer template. Use when drafting a reply to a concrete review comment, when the user quotes a reviewer ("the reviewer says...", "R2 complains..."), asks which strategy fits a concern, or asks for a rebuttal reply template. Complements write-rebuttal (overall process and tactics); this skill picks the response strategy per concern.

  • In this repositoryOne other file it gives its agents.

The file

Files
SKILL.md

NLP Rebuttal Scenario Playbook

Translated and adapted from MLNLP-World/Paper-Rebuttal-Tips. 28 recurring reviewer-concern scenarios. Each has four parts: the concern, a bad answer that backfires, a recommended answer template, and the takeaway.

Good rebuttal = Respect + Evidence + Clarity.

How to use

  1. Classify each reviewer comment with the router below.
  2. Load ONLY the reference file(s) for the matched tips.
  3. Adapt the recommended-answer template: replace every placeholder (XXX, A/B/C, Table X) with the paper's real content and freshly computed numbers. Never ship a template verbatim.
  4. For overall response structure, ordering, and tone, use the write-rebuttal skill (itemize → brain-dump → draft → revise, the 18 tactics, the neutral-third-party test). These two skills compose: that one shapes the whole response, this one shapes each answer.

Concern router

Reviewer concern sounds likeTipReference
"Too complex", "bag of tricks", "which component matters?"1innovation-theory.md
"Not novel", "similar to prior work A"2innovation-theory.md
"Just a combination of existing techniques"3innovation-theory.md
"Contributions unclear"4innovation-theory.md
"Motivation unclear", "why is this problem important?"5innovation-theory.md
"No theoretical analysis", "why does it work?"6innovation-theory.md
"Limitations discussion is superficial"7innovation-theory.md
"Related work missing/insufficient"8communication-writing.md
"Writing/notation unclear"9communication-writing.md
Reviewer misunderstood the method10communication-writing.md
Vague, low-quality negative review11communication-writing.md
Tempted to reply "we will add..."12communication-writing.md
"Missing/weak baselines"13experiments-evidence.md
"Improvements are marginal"14experiments-evidence.md
"Unfair experimental setup"15experiments-evidence.md
"Missing ablations"16experiments-evidence.md
"Too much computational overhead"17experiments-evidence.md
Asked for experiments too large for the rebuttal window18experiments-evidence.md
"Dataset too small"19experiments-evidence.md
"Generalization not shown" (few datasets/models/tasks)20experiments-evidence.md
"No variance / significance / seeds"21experiments-evidence.md
"Possible train/test leakage or contamination"22experiments-evidence.md
"Hyperparameter sensitivity?" ("why k=40?")23experiments-evidence.md
"Wrong/missing evaluation metrics"24experiments-evidence.md
"No human evaluation"25experiments-evidence.md
"Intermediate outputs never evaluated directly"26experiments-evidence.md
Claims "continual/online" but experiments are one-shot offline27experiments-evidence.md
"No code, seeds, or hyperparameters — not reproducible"28experiments-evidence.md

A single comment often maps to several tips (e.g. "marginal gains and no significance testing" = 14 + 21). Load all matches and merge their strategies into one answer.

Cross-cutting rules (from the 28 scenarios)

  • Act, don't promise (Tip 12): run the number/analysis now and put it in the rebuttal. "We will add X in the revision" alone convinces nobody.
  • Never blame the reviewer (Tips 4, 9, 10): if they misread, the fix is a clarification plus a pointer to the line, stated neutrally.
  • Answer head-on (Tips 2, 3): name exactly where the difference or novelty lies — motivation, mechanism, role — not just that it exists.
  • Evidence over adjectives (Tips 13-28): every disputed claim gets a table, an ablation, a test, or an honest statement of infeasibility with a scaled-down proxy result (Tip 18).
  • Concede real weaknesses gracefully (Tips 7, 14, 19): bound the claim, show the trend, explain what the paper still establishes.

Anti-patterns (never do)

  • Asserting all components are necessary without per-component ablation (Tip 1).
  • "We are the first to apply X to Y" as the whole novelty defense (Tip 3).
  • Repeating the introduction as the answer to a motivation question (Tip 5).
  • "Experiments show it works" as the answer to a theory question (Tip 6).
  • Calling the setup fair without matching compute/tuning budgets (Tip 15).
  • Dismissing a metric request instead of adding the metric (Tip 24).

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.