deweight-anchor-rewrite

スター0

フォーク0

更新日2026年5月6日 06:40

学术降重的锚点约束改写门。只根据 Semantic Anchor、Pattern Coverage Matrix 和真人实践卡绑定处方改写 suspect 段；无绑定不改写。触发关键词：按锚点改写、处方式改写、anchor-bound rewrite、去 AI 改写、论文段落改写。

インストール

Codex または Claude でインストールこの Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。

ソース

kkunkunya

kkunkunya/academic-deweight-suite

GitHub リポジトリを開く Creator のリポジトリを見る

ダウンロード

Deweight Anchor Rewrite

Purpose

改写目标是降低 AI 模式残留，不是润色成统一顺滑风格。所有改写必须追溯到：

paragraph_id -> hit_modes -> practice_cards -> rewrite_recipe

Required Inputs

Original paragraph text.
Semantic Anchor.
Human Anchor Binding prescription.
Rewrite Prescription following ../../_shared/references/rewrite-prescription-rubric.md.
Integrity Snapshot for immutable token restore.

Hard Gate

If a suspect paragraph has no valid human-anchor binding, return:

blocked: missing_human_anchor_binding

Do not rewrite.

If the binding has no human_anchor_evidence.quote_exact, return:

blocked: missing_human_quote_evidence

Rewrite Rules

Preserve language of the input.
Preserve claim polarity and evidence strength.
Preserve citation keys, author/year mentions, numbers, units, formulas, table/figure labels.
Change structure before vocabulary.
Read human_anchor_evidence.quote_exact before rewriting; imitate the quoted author's writing move, not the surface wording.
Use observed_human_move, transferable_structure, and adaptation_plan as the concrete operation.
Do not copy quote_exact wording, examples, entities, data, or topic-specific phrasing into the user's rewrite.
Do not copy OCR/MinerU noise from practice cards.
Do not add first-person experience, new examples, new data, or fabricated sources.
Do not use SQL/database retrieval results as authority by themselves; retrieved cards are candidates and still need real practice-card binding.

Output Contract

For each changed paragraph:

paragraph_id:
before_modes:
practice_cards_used:
human_quote_used:
human_move_transferred:
rewrite_recipe_applied:
rewritten_text:
integrity_notes:

If report=rewrite_only, output only the final rewritten text after verification.

The rewrite prescription consumed by this skill should pass:

python3 ../../_shared/scripts/check_rewrite_prescription.py <rewrite-prescription.md>

このリポジトリの他の Skills

同じリポジトリ

deweight-coverage-scan

kkunkunya/academic-deweight-suite

学术降重的全量 AI 模式扫描门。长文必须启用 subagent 并行隔离扫描，生成 Pattern Coverage Matrix，确保每段都有 clean/suspect/format-exempt 状态和 A-code 证据。触发关键词：长文扫描、模式门控、Pattern Coverage Matrix、A-code 扫描、反漏检、母骨架诊断、subagent 并行扫描。

2026-05-060

deweight-human-anchor-bind

kkunkunya/academic-deweight-suite

学术降重的真人写作实践绑定门。用于把每个 suspect 段落绑定到 A-code、写作功能和具体真人实践卡，生成可执行改写处方。触发关键词：真人写法锚点、实践卡绑定、anchor binding、rewrite prescription、A-code 锚点。

2026-05-060

deweight-intake-anchor

kkunkunya/academic-deweight-suite

学术降重的输入锚定门。用于生成 Integrity Snapshot 与 Semantic Anchor，锁定术语、引用、数字、公式、核心命题、段落功能和上下文依赖。触发关键词：文意锚定、Integrity Snapshot、Semantic Anchor、降重前锚定、不可变 token。

2026-05-060

deweight-verify-gate

kkunkunya/academic-deweight-suite

学术降重的残留验收门。用于改写后复扫原 A-code 是否下降、新 A-code 是否冒出、Semantic Anchor 是否漂移、不可变 token 是否漂移。触发关键词：残留验收、Residual AI Gate、改后复扫、文意漂移检查、降重验证。

2026-05-060

orchestrator

kkunkunya/academic-deweight-suite

学术降重主编排器。用于论文、学位论文、基金文本、技术报告的去 AI 味/降重流程；按短文本或长文路线强制执行文意锚定、全量模式扫描、真人写作实践卡绑定、锚点约束改写和残留验收。触发关键词：学术降重、论文降重、去 AI 味、AIGC 检测、academic deweight、母骨架诊断、长文扫描、真人写法锚点。不用于个人口语化改写（那是 deweight-suite:ai-deweight-personal）。

2026-05-060

name	deweight-anchor-rewrite
description	学术降重的锚点约束改写门。只根据 Semantic Anchor、Pattern Coverage Matrix 和真人实践卡绑定处方改写 suspect 段；无绑定不改写。触发关键词：按锚点改写、处方式改写、anchor-bound rewrite、去 AI 改写、论文段落改写。

Deweight Anchor Rewrite

Purpose

改写目标是降低 AI 模式残留，不是润色成统一顺滑风格。所有改写必须追溯到：

paragraph_id -> hit_modes -> practice_cards -> rewrite_recipe

Required Inputs

Original paragraph text.
Semantic Anchor.
Human Anchor Binding prescription.
Rewrite Prescription following ../../_shared/references/rewrite-prescription-rubric.md.
Integrity Snapshot for immutable token restore.

Hard Gate

If a suspect paragraph has no valid human-anchor binding, return:

blocked: missing_human_anchor_binding

Do not rewrite.

If the binding has no human_anchor_evidence.quote_exact, return:

blocked: missing_human_quote_evidence

Rewrite Rules

Preserve language of the input.
Preserve claim polarity and evidence strength.
Preserve citation keys, author/year mentions, numbers, units, formulas, table/figure labels.
Change structure before vocabulary.
Read human_anchor_evidence.quote_exact before rewriting; imitate the quoted author's writing move, not the surface wording.
Use observed_human_move, transferable_structure, and adaptation_plan as the concrete operation.
Do not copy quote_exact wording, examples, entities, data, or topic-specific phrasing into the user's rewrite.
Do not copy OCR/MinerU noise from practice cards.
Do not add first-person experience, new examples, new data, or fabricated sources.
Do not use SQL/database retrieval results as authority by themselves; retrieved cards are candidates and still need real practice-card binding.

Output Contract

For each changed paragraph:

paragraph_id:
before_modes:
practice_cards_used:
human_quote_used:
human_move_transferred:
rewrite_recipe_applied:
rewritten_text:
integrity_notes:

If report=rewrite_only, output only the final rewritten text after verification.

The rewrite prescription consumed by this skill should pass:

python3 ../../_shared/scripts/check_rewrite_prescription.py <rewrite-prescription.md>