# Academic De AI Polishing

> Use when polishing stable paper, review, thesis, rebuttal, or other scholarly prose that reads mechanically, over-smoothed, obviously AI-assisted, or contaminated by internal project notes, operation logs, process records, TODOs, handoff language, or draft-management traces in the manuscript body. Best after content, structure, evidence, citations, and claim strength are stable, for section-by-section rewriting that preserves meaning, citations, and technical rigor while reducing templated transitions, repetitive cadence, formulaic academic phrasing, and non-manuscript operational residue. Also use when a paragraph needs section-appropriate rhetorical calibration between正文机理密度 and摘要/引言/过渡 compactness. Chinese triggers: 论文降AI痕迹, 去AI味, 去模板化润色, 降低机械感, 学术润色, 最终润色, AIGC痕迹评估, 正文机理升级, 摘要过渡压缩, 内部项目记录, 操作记录痕迹, 主文日志化.

- Skill: `cuimuxuan/academic-de-ai-polishing` (Agent Skill, multi-file: 12 files)
- Install (CLI): `npx skillmds@latest add cuimuxuan/academic-de-ai-polishing`
- Raw SKILL.md: https://api.skillmd.com/api/skills/cuimuxuan/academic-de-ai-polishing/raw
- Safety review: pending (external: skill-scanner PASS, skillspector PASS)
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: CuiMuxuan (https://skillmd.com/u/cuimuxuan)
- Updated: 2026-09-22
- Page: https://skillmd.com/skills/cuimuxuan/academic-de-ai-polishing

---


# Academic De-AI Polishing

Use this skill to reduce mechanical or AI-like academic prose while preserving the core argument, evidence balance, citation logic, terminology, technical scope, and section-appropriate rhetorical mode.

## Boundaries

Do:

- Rewrite local paragraphs or adjacent paragraph blocks.
- Reduce templated transitions, repetitive cadence, vague academic filler, and over-smooth balance.
- Remove or recast internal project notes, operation logs, prompt traces, TODO markers, and draft-management residue.
- Preserve or strengthen evidence-grounded authorial judgment.
- Enforce final-polish boundaries after writing, evidence, and structure are stable.
- Execute confirmed local rewrite tasks from `$revision-control` when the task is language-level and state boundaries are already set.

Do not:

- Fabricate evidence, citations, results, or conclusions.
- Change claim strength beyond what the evidence supports.
- Claim detector percentages or guarantee AIGC-score outcomes.
- Regenerate whole sections unless the user explicitly asks.
- Use polish to bypass missing evidence, unresolved `LIT_GAP` markers, or unstable manuscript structure.
- Act as a language-style review skill; route diagnosis-only requests to `$language-style-review`.
- Maintain sentence-level revision status, object libraries, latest drafts, or round logs; route those tasks to `$revision-control`.
- Handle literature补充, structural reorganization, or large-scale manuscript rewriting as the owner skill.

## Core Rules

1. Preserve technical meaning, evidence balance, citation logic, section purpose, and terminology.
2. Fix argument, evidence scope, and claim precision before fixing cadence.
3. Prefer paragraph-level rewrites over whole-section regeneration.
4. Do not make prose more confident than the evidence permits.
5. Keep Chinese, English, and bilingual terminology stable unless the user asks for translation.
6. Move unresolved author decisions to `User-review items` instead of hiding them in polished prose.
7. Treat accurate, direct, concise, understandable writing as the baseline; style elevation is optional and must stay evidence-safe.
8. In main text and rebuttal prose, do not polish unsupported claims into final language; request support or soften the claim.
9. Use the shared cross-disciplinary language review gate before finalizing formal body-text or rebuttal rewrites with abstract terms, local labels, operation-record residue, unclear sentence purpose, proxy wording, or bilingual inconsistency.
10. When the target is a正文机理段, allow denser conceptual compression and explicit logic chaining if the evidence and terminology support it; when the target is摘要、引言、结论 or a long transition, keep the chain compact and reduce connector stacking.
11. When called by `$revision-control`, only return the confirmed local rewrite and hand state updates back to `$revision-control`.

## Pre-Edit Gate

Before rewriting, confirm or infer:

```text
Source text present:
Section type:
Paragraph role:
Fixed claims and citations:
Claim anchors, if available:
Allowed edit intensity:
Do-not-change terms:
```

If source text is missing, ask for it. If citation or evidence boundaries are unclear, keep the rewrite conservative and mark author verification needs.

## Argument-First Gate

Classify each target paragraph:

| State | Action |
|---|---|
| `sound_argument_mechanical_style` | Rewrite for cadence, vocabulary, transitions, and authorial voice. |
| `overclaiming_or_missing_evidence` | Flag the claim problem first, then soften or request evidence before polishing. |
| `unclear_section_job` | Ask for section role or propose a local role before rewriting. |
| `citation_boundary_unclear` | Preserve citations conservatively and mark claims that need author verification. |
| `internal_project_or_operation_trace` | Remove or recast workflow residue using `manuscript-residue-cleanup.md`. |

## Workflow

1. Identify medium- and high-risk paragraphs.
2. Diagnose the dominant risk: connector overload, repetitive cadence, templated dialectic, vague vocabulary, missing judgment, overclaiming, or workflow residue.
3. Use [final-polish-boundary.md](references/final-polish-boundary.md) before broad final polishing or when the request may hide evidence, structure, or claim-strength work.
4. Use [high-risk-patterns.md](references/high-risk-patterns.md) for pattern diagnosis and [rewrite-tactics.md](references/rewrite-tactics.md) for revision tactics.
5. Use [section-playbooks.md](references/section-playbooks.md) when the rewrite problem is section-dependent.
6. Use [manuscript-residue-cleanup.md](references/manuscript-residue-cleanup.md) when main text contains project, operation, prompt, or file-management traces.
7. Use [main-text-and-rebuttal-claim-support-gate.md](../../shared/main-text-and-rebuttal-claim-support-gate.md) when final main-text or rebuttal wording contains factual, boundary, subjective, contribution, or limitation claims.
8. Route diagnosis-only language review, terminology review, and strict sentence-by-sentence checking to `$language-style-review`; use [cross-disciplinary-language-review-gate.md](../../shared/cross-disciplinary-language-review-gate.md) only as a local safeguard during rewriting.
9. After each rewrite, verify meaning, citations, terminology, claim anchors, support wording, and technical scope.

## Material Request

Before assessment or rewriting, ask for missing materials that would materially improve quality:

```text
Required or high-value materials:
Missing materials:
Can proceed without them: yes/no
Fallback if unavailable:
```

Ask for source text, target examples, section type, paragraph role, fixed claims, fixed citations, do-not-change terms, allowed edit intensity, and whether the user wants diagnosis only or rewrite.

## Reference

Read [edit-contract-template.md](references/edit-contract-template.md) when the edit boundary is unclear.

Read [final-polish-boundary.md](references/final-polish-boundary.md) when deciding whether the manuscript is ready for final polish or whether the task should route back to writing, research verification, or post-draft review.

Read [high-risk-patterns.md](references/high-risk-patterns.md) when diagnosing mechanical or AI-shaped sentence patterns.

Read [rewrite-tactics.md](references/rewrite-tactics.md) for paragraph-level rewrite tactics, target-example calibration, and output patterns.

Read [before-after-examples.md](references/before-after-examples.md) when the user wants examples or when calibration is needed.

Read [section-playbooks.md](references/section-playbooks.md) for abstract, introduction, technology, discussion, review, or thesis-section handling.

Read [project-notes-from-this-paper.md](references/project-notes-from-this-paper.md) for long review-paper polishing sequences.

Read [manuscript-residue-cleanup.md](references/manuscript-residue-cleanup.md) for internal project notes, operation logs, prompt traces, file-management language, handoff language, TODO markers, or draft residue.

Read [main-text-and-rebuttal-claim-support-gate.md](../../shared/main-text-and-rebuttal-claim-support-gate.md) before polishing unsupported main-text or rebuttal claims into final prose, and when replacing generic "evidence" wording with a more precise support term.

Read [cross-disciplinary-language-review-gate.md](../../shared/cross-disciplinary-language-review-gate.md) before strict language review or final polishing of formal body text or rebuttal prose that may contain abstract wording, undefined local terms, operation-record traces, unclear sentence function, unsupported proxies, citation-distance problems, or bilingual claim-strength drift.

Read [revision-control-contract.md](../../shared/revision-control-contract.md) when a rewrite task arrives from `$revision-control`.

Read [ai-trace-assessment.md](references/ai-trace-assessment.md) when the user asks for AI-trace or AIGC-risk assessment.