Sci-Polish — Academic Paper Polishing & De-AI

Two-stage academic paper polishing: AI trace reduction + 8-dimension quality improvement.

Hard Rules (NEVER violate)

1. NEVER modify: LaTeX commands, math environments ($...$, \[...\], equation, align, etc.), \cite{}, \ref{}, \label{}, algorithm blocks, code listings, figure/table environments (only caption text is editable)
2. NEVER fabricate: data, citations, experimental results, baselines, metrics, author names
3. NEVER remove: author's technical claims without explicit permission
4. NEVER introduce: Tier 1 AI words (see references/word-choice-anti-ai.md) during Phase B edits
5. Mark as [PENDING VERIFICATION]: any claim you cannot verify from the text
6. Preserve meaning: the exact semantic content of every technical statement must be unchanged
7. Output format: revised text + polish report. No preamble or trailing changelog unless user asks
8. When uncertain about domain terminology: keep the original, flag for user review

Execution Modes

Step 0: Input Processing

1. Identify input format:

- PDF → extract visible text (preserve section structure) - LaTeX → extract visible paragraphs (protect all commands) - Markdown / plain text → use directly

1. Identify protection zones: math environments, citations, labels, algorithm blocks, code blocks, figure/table environments
2. Ask target journal (optional): affects Dimension 8 scoring. If not specified, use general academic standards.
3. Confirm execution mode: full / deai-only / polish-only / polish-select

Step 1: Phase A — De-AI Editing

Reference files:

• references/ai-detection-patterns.md — pattern library for scanning
• references/de-ai-strategies.md — rewriting strategies
• references/word-choice-anti-ai.md — word replacement table

1.1 Scan

Scan text against ai-detection-patterns.md. Categorize findings:

• High priority: Tier 1 words, template phrases, burstiness deficit (SD < 8), RLHF fingerprints
• Medium priority: Tier 2 threshold violations, structural symmetry, transition stacking
• Low priority: Tier 3 context-dependent, punctuation fingerprints, minor style uniformity

1.2 Rewrite (Priority Order)

Follow the strategy priority from de-ai-strategies.md:

a. Burstiness enforcement — restructure sentence lengths (Strategy 1)
   Target: SD > 8, range ≥ 20, mid-band < 50%

b. Vocabulary de-AI — replace Tier 1 words, compress template phrases (Strategy 2)
   Consult word-choice-anti-ai.md for replacements

c. Sentence restructuring — vary syntax patterns (Strategy 3)
   No 3+ consecutive same-structure sentences

d. Concept concretization — add specifics where available (Strategy 4)
   Replace vague claims with data from the paper itself

e. Argumentation enrichment — add depth to claims (Strategy 5)
   Insert boundary conditions, alternative interpretations

f. Perplexity elevation — introduce unpredictability (Strategy 6)
   Unconventional openers, domain-specific minor terms

g. Style disruption — break uniformity (Strategy 7)
   Vary paragraph density and register between sections

1.3 Self-Audit

After rewriting, re-scan output against ai-detection-patterns.md:

• Any remaining High-priority patterns? → fix them
• Burstiness check: SD > 8? Range ≥ 20? Mid-band < 50%?
• Protection zones intact? (LaTeX commands, math, citations unchanged)
• Meaning preserved? (no semantic drift from original)

Step 2: Phase B — 8-Dimension Polish

Reference file: references/polish-dimensions.md

2.1 Score All Dimensions (1-5)

Rate each dimension based on the checkpoints in polish-dimensions.md:

1. Grammar & Syntax
2. Academic Tone
3. Coherence & Flow
4. Conciseness
5. Terminology
6. Structure & Formatting
7. Clarity of Argument
8. Journal Compliance

2.2 Edit Dimensions Scoring ≤ 3

Edit in this order (foundation → logic → style → format):

Dim 1 (Grammar) → Dim 7 (Clarity) → Dim 3 (Coherence) →
Dim 4 (Conciseness) → Dim 2 (Tone) → Dim 5 (Terminology) →
Dim 6 (Structure) → Dim 8 (Journal)

2.3 Preservation Check

During Phase B editing, verify Phase A improvements are maintained:

• Burstiness NOT reduced (sentence length variety preserved)
• No Tier 1 AI words reintroduced
• No template phrases reintroduced
• Punctuation fingerprints still clean (em dash count, etc.)

2.4 Re-Score

Rate edited dimensions again. Record "Before" and "After" scores for the report.

Step 3: Output

3.1 Revised Text

Output the complete polished text with all modifications applied. Format matches the input format (LaTeX → LaTeX, Markdown → Markdown, etc.).

3.2 Polish Report

Generate report using templates/polish-report.md:

• Phase A summary: Number of AI traces detected and fixed, by priority level
• Phase B scorecard: 8 dimensions, before/after scores, key changes per dimension
• Top modifications: Up to 10 most significant before/after comparisons with reasoning
• Pending items: Anything marked [PENDING VERIFICATION] or requiring user decision

Section-Specific Guidance

Abstract

• Phase A focus: Remove "In recent years", overclaiming words, template openings
• Phase B focus: Ensure all 6 elements present (context, gap, approach, insight, results, significance)
• Length: Check against venue limits (typically 150-250 words)

Introduction

• Phase A focus: Break three-part parallel structures, remove "has attracted growing attention"
• Phase B focus: Contributions must be specific and verifiable, not vague

Methods

• Phase A focus: Restore active subjects ("We" not passive-only), vary procedural sentence length
• Phase B focus: Every symbol defined before use, equations motivated and interpreted

Results

• Phase A focus: Replace "clearly demonstrates" with data speaking for itself
• Phase B focus: Every claim backed by table/figure reference, statistical measures included

Discussion

• Phase A focus: Break balanced-hedging patterns, ensure position is taken
• Phase B focus: Limitations acknowledged specifically, future work concrete

Conclusion

• Phase A focus: Remove "In conclusion, this paper presents" opener
• Phase B focus: Brief, no methodology restatement, concrete future work

Voice Calibration (Optional)

If the user provides their previously published papers:

1. Extract author writing profile (sentence length distribution, connector preferences, punctuation habits)
2. Use author's actual patterns as the target instead of generic "human-sounding" defaults
3. Note in report: "Voice calibrated against [paper title]"

This replaces generic de-AI rewrites with author-specific style matching.

Usage Examples

Full polish:

"请润色这篇论文" / "Polish this paper for NeurIPS submission"

De-AI only:

"只帮我降AI率" / "Reduce the AI detection score, don't change anything else"

Selective polish:

"只检查语法和术语一致性" / "Just fix grammar and check terminology consistency"

With voice calibration:

"用我之前发的这篇论文的风格来润色" / "Polish matching my writing style from this published paper"

© License

MIT License — © 2026 Shuo Zhao. All rights reserved.

This skill is part of the AutSciWrite suite.