rebuttal-guidance

Rebuttal Guidance

Overview

This skill delivers executable rebuttal guidance per concern: structured plans, evidence pointers, tone and strategy priorities. It does not produce ready-to-paste rebuttal prose; the author writes the final text.

The core model is conditional persuasion P(Z=1 | X, Y):

• X: reviewer mindset (behavior cluster inferred from review text)
• Y: rebuttal strategy vector (which tactics to emphasize)
• Z: success (score increase or paper accepted)

When to use

• User pastes Weaknesses / Questions from ICLR, NeurIPS, ICML, EMNLP, etc.
• User asks how to respond to a specific reviewer comment.
• User wants a structured plan before writing rebuttal prose.
• User asks which tone or evidence strategy fits a harsh vs. open reviewer.
• User says "reviewer ignored my rebuttal" and wants to recalibrate.

When NOT to use

• User wants full rebuttal text copied verbatim → output guidance first; warn against blind LLM paste.
• User asks to evaluate a research idea → use idea-evaluator.
• User asks for pre-submission paper review → use pre-submission-reviewer.
• No review text provided → ask for it first.

Core procedure

Step 1: Inventory concerns

Extract every numbered Weakness and Question from the pasted review. If the user pasted only one reviewer's block, ask whether other reviewers exist.

Output a short table:

Step 2: Classify each concern

See: references/concern-classification.md

Tag each item as one of: misunderstanding | evidence_gap | novelty | baseline_fairness | scope_claim | writing_clarity | theory | efficiency | other

Step 3: Match reviewer mindset (X)

See: references/mindset-matching-heuristics.md

Score the review on 5–7 behavior signals (openness, severity, constructiveness, specificity, skepticism, harshness, actionability). Map scores to a cluster in references/mindset-library.md.

State: matched cluster name, confidence (high / medium / low), and the specific phrases that triggered the match.

Step 4: Strategy priorities for this mindset (Y)

From the matched cluster in references/mindset-library.md, read:

• Historical up-rate: calibrate how likely a score increase is.
• Rule-stat diffs (success vs. failure %): boost tactics with positive diff.
• Cohen's d effects: highlight dimensions with d > 0.8 in success group.
• Success / failure patterns: concrete do's and don'ts from data.
• Key strategy: one-sentence summary for this cluster.

Cross-check dimensions in references/strategy-dimensions.md.

Step 5: Per-concern guidance

For each concern, emit a guidance block (not final prose):

### R?.W?: [short title]
- **Concern type**: ...
- **Mindset-aware strategy**: ...
- **Recommended structure**: Acknowledge → Response → Evidence → Revision note
- **Clarify from paper?**: yes/no: where to look (section, table, appendix)
- **New experiment needed?**: yes/no: what exactly (no fabricated numbers)
- **Tone note**: ...
- **Template**: A / B / C: see references/response-templates.md
- **Avoid**: ...

Do not fabricate experimental numbers. Mark anything unverified as [TO VERIFY].

Step 6: Integrity gate

Before finalizing, run through references/checklist.md.

Key flags:

• Every W/Q has a guidance block.
• No fabricated numbers.
• Defensive phrases (e.g., "you misunderstood") flagged.
• Overpromise risk flagged for scope_claim and evidence_gap types.

Step 7: Output format

Deliver a ## Rebuttal Guidance section containing:

1. Reviewer mindset summary: cluster name, confidence, identifying phrases.
2. Global strategy priorities: 3–5 bullets from mindset data.
3. Per-concern guidance: Step 5 blocks for each W/Q.
4. Checklist highlights: pass/fail on key integrity items.

Reference map