Imported from Moonlit-Pages/AIGC-Detector-Rewriter-Skill (SKILL.md). Install upstream with npx skills add Moonlit-Pages/AIGC-Detector-Rewriter-Skill. Copyright stays with the author.

AIGC Detector & Rewriter Skill

Description

Detect AI-writing risk patterns in English and Chinese thesis text — English text uses the conservative D1–D17 framework, Chinese text uses the Chinese-text AI-risk module plus shared structure-level checks, and mixed documents are routed by language segment — score paragraph-level and document-level AI risk, and rewrite high-risk sections by reducing shared external-detector-sensitive features such as repeated templates, uniform rhythm, mechanical reporting sequences, and over-smooth academic reasoning. The scoring system is intentionally conservative and calibrated for structural AI-risk signals commonly detected by external AI detection systems in formal master’s thesis writing.

This skill should not treat a low internal score as proof that the text is safe. When a known detector result is available, such as Turnitin AI = 48%, use it as a calibration signal and adjust the internal scoring more strictly.

Core Operating Architecture (READ FIRST — overrides any conflicting rule below)

Multiple failure modes have been observed in real multi-round use, including score regression, structural over-editing, word-count loss, readability damage, detector-comparison error, and unsupported researcher-trace insertion. The rules in this section exist to prevent them and override any conflicting rule anywhere else in this file.

Principle 0 — The goal is a TARGET, not merely "lower than last time"

The operational objective is to reduce traceable AI-writing-risk patterns in user-provided thesis text and to prepare candidate revisions for comparable external re-test. The target threshold (default 30%, or a stricter value the user specifies) is used only as a closed-loop stopping condition when comparable detector evidence exists; the skill must never guarantee, imply, or optimise for a detector outcome independently of academic integrity, meaning preservation, and evidence-based revision. Chapter scores are the primary closed-loop units; the whole-document score is a target only when a comparable whole-document external test is available. "Lower than the previous version" is NOT success. A version that is the lowest so far but still above the active target is NOT finished — it is work in progress.

Authorized-Use Boundary: this skill may be used only on thesis text the user owns, authored, co-authored, or is explicitly authorized to edit. It supports clarity, academic integrity, and detector-informed risk diagnosis; it must NOT be used to misrepresent authorship, hide prohibited AI use, fabricate research contribution, or bypass an institution's disclosure or assessment rules.

T19/T29 Master Boundary Rule: T19 is relocation or light integration of researcher-side material that already exists elsewhere in the thesis; T19 never creates new interpretive content. T29 is evidence-based reconstruction; T29 may add a new interpretive sentence only when a completed Source Trace block proves it is grounded in existing thesis materials, codes, quotations, methods, findings, or cited literature. For qualitative passages, any new interpretive sentence must use T29, never T19. If a source trace cannot be established, write "Author confirmation needed" instead of adding the sentence.

Strict Target Warning: if the user sets a target below 30% (especially below 15%), warn that the target may be unrealistic depending on detector behaviour, thesis genre, standardised methodology language, and required academic terminology; the skill can attempt risk-pattern reduction but cannot guarantee the target; and academic integrity, protected content, readability, and meaning preservation override the numeric target. A stricter user target changes the target threshold ONLY — it never relaxes any integrity, minimal-edit, no-fabrication, word-count, or quality-gate rule.

Definition of done:

• A chapter is DONE only when its measured external score is < 30% (or the user's target).
• The whole task is DONE only when every externally measured target unit is below target. If the external detector supports only chapter-level testing, the task may be DONE once every chapter is below target; if a comparable whole-document score is available, the whole document must also be below target.
• Until then, the skill must keep proposing the next rewrite round. It must never present an above-target version as the final result or imply that no further reduction is possible.

Never-stop-early rule:

• Reaching a new lowest score does not end the work; it only updates the baseline. If that lowest score is still ≥ 30%, immediately plan the next round on the highest-remaining chapter.
• The legitimate reasons to pause are: (a) all measurable target units are below target; (b) the user asks to stop; (c) the full escalation ladder (Principle 5) has been exhausted and documented on a chapter AND the user has been told which specific structural constraint is blocking further reduction; (d) Rewrite Fatigue Detection is triggered; (e) no comparable external re-test is available; or (f) continuing would risk changing protected academic content or damaging readability. Rewrite Fatigue Detection overrides the never-stop-above-target rule: when fatigue triggers, restore the lowest-ever comparable version, report the next untried safe option, and do not run another round until the user chooses to continue.
• "I cannot reduce it further" may only be said after every rung of the Principle 5 escalation ladder has been tried and documented for that chapter.

Principle 1 — Closed loop, one chapter per round

This skill does NOT rewrite a whole thesis in one response. Internal D1–D17 scoring cannot reliably predict the external detector result, so the only trusted verdict is a real re-test.

The mandatory loop is:

pick the single highest-scoring chapter that is still at or above the target (default 30%)
  → rewrite ONLY that chapter, using minimal edits at the current escalation rung
  → output the rewritten chapter + evidence
  → STOP and ask the user to re-test with CNKI / Turnitin
  → if the new score is BELOW target: lock it as done, move to the next-highest chapter
  → if the new score is lower than before but STILL above target: keep this version as the new baseline, mark the chapter PROGRESS (not DONE), AND plan the next round on the SAME chapter, moving UP one escalation rung (Principle 5). "Plan" means describe the next round; do NOT execute the next rewrite in the same response. Wait until the user supplies a comparable external re-test of this version (or the chapter text/version to revise) or explicitly asks to continue, per the mandatory re-test stop.
  → if the new score stayed the same or went UP: roll back to the chapter's lowest-ever version, then move UP one escalation rung and try a DIFFERENT technique on the same chapter
  → repeat until the chapter is below target or the escalation ladder is exhausted

Always tell the user explicitly: "Rewrite one chapter, re-test it, and continue. We keep going on each chapter until it is below the active target threshold, not just lower than before." The active target is the user-specified threshold when provided; otherwise it defaults to below 30%.

Principle 2 — Minimal edit only, never regenerate

The biggest cause of score regression (e.g. a chapter rising from 36% to 100%) is letting the model re-generate a paragraph as fluent new prose. Fluent AI-generated academic text carries strong AI features by construction.

Rules:

• Rewrite by editing the EXISTING sentences: substitute individual words, reorder clauses, split long sentences, recast voice/structure, and relocate or integrate researcher-trace material drawn from existing content.
• Every sentence in the output must be traceable to a specific original sentence (or be a permitted researcher-trace/breakpoint sentence based on existing content).
• Preserve as much of the original human wording as possible — original human phrasing is what keeps the score down.
• Whole-paragraph or whole-section regeneration is FORBIDDEN, even for A-class / P0 paragraphs.
• Back-translation is FORBIDDEN. Never translate text to another language and back (e.g. Chinese → English → rewrite → Chinese) to reduce AI features. Round-tripping through translation is whole-paragraph regeneration in disguise: it discards the original human wording and emits fresh machine-distributed text — the exact failure mode that drove a chapter from 36% to 100%. Edit the existing sentences in place instead.
• Do NOT delete content or merge sentences to reduce AI features. Splitting one sentence into two is allowed; merging two sentences into one is not allowed AS A ROUTINE AI-RISK TECHNIQUE (it loses length and is rarely necessary). Exception: a merge IS permitted to REPAIR a fragment, an accidentally split clause, or a thin/over-conversational sentence created during revision (Gate H); the repaired sentence must preserve the original meaning and must not reduce the chapter's substantive content. See Principle 4.
• "Aggressive Structural Rewrite" means aggressive REORGANISATION of existing sentences (reordering, re-segmenting, changing entry point), never aggressive REWRITING into new prose.

Priority note: the principles are numbered in order (P1–P6) and the workflow Phases run in order (Phase 1 → 1.5 → 1.6 → 1.6A → 1.7 → 2 → 2b → 3 → 4 → 4.5 → 4.5A → 5 → 5.5 → 6 → 7 → 8 → 9). Numeric order is for reference only; it is NOT the conflict-resolution order. When two rules ever appear to conflict, apply this fixed priority instead: protect data/citations/quotes/hedges > preserve word count > readability soft-limits > AI-risk reduction; and external re-test result > internal risk estimate.

Conflict Resolution Hierarchy (full order, highest first): (1) academic integrity and protected elements; (2) user-specified scope and target; (3) Core Operating Architecture and Principles 0–9; (4) language-route rules (English D1–D17 vs Chinese route); (5) task-mode and staged-execution rules; (6) Phase gates A–J; (7) reference modules and technique files; (8) examples and prompts. Examples never override mandatory rules. Within this order, the data/word-count/readability/AI-risk and external-vs-internal sub-priorities above still apply.

Canonical Output Authority: SKILL.md is the only authority for workflow-level output structure (Phase order, target lock, ledger, Phase 7 Gate A–J, Phase 8/9 summary, external re-test stop, DOCX Output Gate). Reference files may provide paragraph-level evidence templates, examples, and technique-specific subfields only; if any reference output template conflicts with SKILL.md, SKILL.md overrides it.

Minimum Execution Map (anti-phase-skipping; full definitions remain in their own sections):

Route / Mode	Required phases (cannot be silently skipped)	Required visible evidence
Quick Mode	Protected-element check → single-paragraph diagnosis → rewrite → Gate A–J summary	Scope statement; hit dimensions or Chinese-route groups; techniques; one-line Gate A–J; "no chapter/document score" statement
Working Mode (English route)	1 → 1.5 → 1.6/1.6A → 1.7 → 2 (full tables for target/highlighted/cluster/≥40% paragraphs) → 3 → 4 → 4.5/4.5A → 5/5.5 → 6 (scope-aware) → 7 → 8 → 9	Full evidence for target paragraphs; compact skip records for non-target low-risk paragraphs; ledger; re-test stop
Working Mode (Chinese route)	Same phase chain, with Phase 2 using Chinese-route fields instead of D1–D17 tables	Chinese-route evidence template; shared structure checks; chinese_paragraph_risk_estimate
Full Staged Mode / Thesis-Evidence Audit	All phases, staged by batch; Stage 1 diagnosis and planning only → Stage 2 rewrite with execution evidence → Stage 3 Gate A–J quality check → Stage 4 scope-aware final output → Stage 5 mandatory re-test stop	Stage outputs incl. appropriate ledger, Master Overview, scope-appropriate risk index, DOCX gate when applicable
Skill-Audit Mode	No thesis phases; no D-scoring	Issue list in the required (or user-provided) audit format

Principle 3 — Red lines, target lock, and the version ledger

• Red line (never regress): Each chapter's lowest-ever measured external score is a hard ceiling for that chapter. A rewrite that scores higher than the chapter's lowest-ever version is discarded and the lowest-ever version is restored.
• Target, not ceiling: The red line prevents regression; it does NOT define success. Success is reaching below the target threshold (default 30%). Keep working a chapter until it is below target, even after it reaches a new lowest score.
• Protection floor = the target: Only a chapter already below the target threshold (default 30%) is protected and left alone. A chapter between 30% and 40% is NOT protected and must keep being rewritten — earlier versions of this skill wrongly froze chapters below 40%, which caused premature stopping above the 30% goal. Do not protect any chapter that is still at or above 30%.
• Target lock: Each round may only touch ONE chapter by default: the highest current measured chapter that is still at or above the target. Everything below target is left untouched. Bounded-unit exception: if an external report highlights only a bounded subsection or a fragment cluster (e.g. one flagged sub-section of the introduction), the round MAY target that bounded unit instead of the whole chapter, while still tracking it under that chapter's ledger.
• Escalation, not surrender: If a chapter changes by less than 5 percentage points in a round but is still above target, do NOT stop and do NOT merely "ask the user to confirm". Move up one rung on the escalation ladder (Principle 5) and try a stronger, still-safe technique on the same chapter next round.

Principle 4 — Word-count preservation (never shrink the paper)

Compression and merging are forbidden as AI-reduction methods. A thesis usually has a minimum word-count requirement, and shrinking it is a serious failure even if the AI score drops.

Rules:

• Every rewritten paragraph must have word count ≥ the original paragraph (one-to-one rewrite), with the acceptable range being the original length up to about +15%. If an over-cap original paragraph is split into two or more descendant paragraphs at safe logical boundaries, Gate F is assessed on the COMBINED word count of all descendants against the original parent paragraph (record Parent Paragraph ID → New Paragraph IDs); the combined count must be ≥ the original parent.
• Over-cap exception (resolves the conflict between "never shrink" and the 200-word / 12-sentence readability cap): the +0% to +15% growth target applies only when the original paragraph is WITHIN the readability cap. If an original paragraph already exceeds 200 words or 12 sentences, do NOT force further expansion. In that case preserve every piece of substantive content, meaning, evidence, citation, variable, datum, and conclusion; do not add filler merely to grow the count; keep the revised paragraph no longer than the original unless a source-supported clarification genuinely requires it; and prefer splitting the over-long paragraph at a natural logical break (this raises sentence count and readability without losing content) over leaving one mega-paragraph. Record Gate F as "Pass — no substantive shrinkage; original already exceeded the readability cap." This exception NEVER licenses removing or condensing substantive content; it only lifts the artificial growth target for paragraphs that are already too long.
• Content deletion vs connector removal (resolves a common deadlock): deleting substantive content — data, reasoning, citations, claims, examples, limitations, findings — is FORBIDDEN. Removing a low-information connective or empty formulaic opener ("Furthermore,", "It is worth noting that", "综上所述") IS allowed, but only when (a) the logical relation stays clear without it and (b) the words removed are compensated within the same paragraph by concrete detail already in the thesis. Connector removal with same-paragraph compensation is NOT content deletion and does not violate the no-shrink rule.
• AI features must be reduced by CHANGING how ideas are expressed (reorder, recast, vary rhythm, differentiate structure, integrate researcher reasoning already present), never by REMOVING ideas or condensing them.
• "Compression", "condensing", "merging templates", "shortening", "trimming", and "removing repeated explanation" are all FORBIDDEN as length-reducing operations. Where older techniques (e.g. T17, T18) say "compress", reinterpret them as "rewrite the repeated template into varied, equally-long-or-longer prose" — break the pattern by restructuring and adding specificity, not by cutting words.
• If breaking a repeated template would remove words, you must compensate by expanding elsewhere in the same paragraph (e.g. add concrete detail, a researcher-view judgment sentence, or a process narration sentence drawn from existing content) so net length does not fall.
• Splitting a long sentence into two shorter sentences is encouraged (it varies rhythm AND tends to add words). Merging sentences is not allowed as a routine AI-risk technique, but IS allowed to repair a fragment or thin sentence (the repair exception in Principle 4 / Gate H).
• Phase 7 Gate F enforces this: any rewritten paragraph shorter than its original FAILS the gate and must be redone.
• Word-count preservation must NOT be achieved by generic filler, duplicated reasoning, empty transition sentences, or unsupported elaboration. Any added wording must preserve or clarify existing content and must pass Gate J's low-information-filler review. Reaching length by padding is itself an AI signal and a quality defect.
• Report the before/after word count for every rewritten paragraph (see Phase 5.5 and Phase 8).

Principle 5 — Escalation ladder (how to keep reducing a stubborn chapter)

When a chapter is still above target, escalate through these rungs in order. Each rung is progressively more structural but every rung still obeys minimal-edit (Principle 2) and word-count preservation (Principle 4). Never jump straight to a high rung; never regenerate prose; never compress.

Rung 1 — Sentence-level: vary openings, transitions, sentence length, lexical choices (T1–T13).
Rung 2 — Rhythm & cadence: break uniform sentence-length patterns; split long sentences; vary clause order within sentences (T11, Pass 2).
Rung 3 — Paragraph architecture: change each paragraph's entry point and internal order; differentiate neighbouring paragraphs that share a template (T14, T15).
Rung 4 — Reporting-sequence & template breaking: reorder reporting steps; de-template methodology/variable/result patterns (T17, T18 — length-preserving versions).
Stop condition (centralised): escalation stops only by marking the unit PAUSED or BLOCKED, and only when academic integrity or protected content would be put at risk, readability damage cannot be avoided, external evidence is non-comparable, Rewrite Fatigue triggers, the user asks to stop, or the full six-rung ladder has been exhausted and documented. Escalation never stops by silently marking an above-target unit DONE.

Rung 5 — Source-extracted researcher-trace integration: relocate or lightly integrate more researcher-view judgment, process narration, scope-narrowing, boundary, and existing-data-anchor material already present in the thesis across the chapter (T19), as a previously observed potentially useful reducer (effectiveness not guaranteed; confirm by comparable external re-test). Never generate new researcher voice. "Integration" here means relocation of existing material, NOT saturation: use it sparingly, at most one relocated researcher-trace sentence per high-risk paragraph, and never insert researcher-trace sentences mechanically across consecutive paragraphs (repeated mechanical insertion is itself an AI-risk pattern).
Rung 6 — Functional reorganisation: change what each paragraph DOES (e.g. turn a flat "definition → source → result" report into a problem-driven or contrast-driven passage) by reordering and recasting the sentences WITHIN that paragraph. This never means shuffling the order of paragraphs across the section or changing the chapter's argument flow (see Principle 7). Still never regenerate prose; keep word count and the paragraph's position unchanged.

Rules:

• Track the current rung per chapter in the ledger.
• After each re-test that is still above target, advance one rung (or, if the last rung produced a clear drop, repeat the same rung more thoroughly before advancing).
• Only after Rung 6 has been applied thoroughly and the chapter still will not drop below target may the skill tell the user that a specific, named structural constraint is blocking further reduction (e.g. an unavoidable standardized methodology section) — and even then it should offer the best available below-baseline version and ask whether to continue trying.
• Escalating rungs must never violate Principles 2 and 4 (no regeneration, no shrinking).

Rewrite Fatigue Detection (mandatory stop condition): Even before Rung 6 is exhausted, declare rewrite fatigue and pause if ANY of these is true for a chapter:

• 3 or more complete rounds (rewrite + external re-test) have been done on this chapter (this is a hard stop checkpoint — the user may explicitly authorize ONE additional controlled round, but the skill must never silently continue past it); OR
• the change in measured external score across the last two rounds is < 3 percentage points (in either direction); OR
• a rewrite increased the score (already handled by Score Regression Guard — fatigue is a wider net).

On fatigue, do NOT keep grinding the same chapter. Take this sequence:

1. Roll back to the chapter's lowest-ever measured version.
2. Look at the "Tried techniques on this rung" column in the ledger (see below). Try ONE clearly untried technique appropriate to the current rung.
3. If no untried technique remains for the current rung, advance one rung.
4. If Rung 6 is reached and still fatigued, STOP this chapter, report the situation honestly to the user, and recommend external editing rather than another model-driven round. Never imply that further mechanical rewriting will help when the data says it will not.

Principle 6 — Preserve structure and readability (no merged mega-paragraphs, no displaced headings)

Reducing the AI score must never make the paper read worse. Two failure modes are forbidden:

(a) Displaced or absorbed headings. Section and sub-section headings such as "1.1 Research Background", "Chapter 2 Theoretical Foundations", "5.2.1 Procedural Control Methods", as well as figure/table captions and numbered/bulleted list markers, are HARD STRUCTURAL BOUNDARIES.

• Never merge body text across a heading.
• Never pull a heading into a body paragraph or let a paragraph run straight into the next heading.
• Never move a heading, renumber it, or strand it at the end of a paragraph.
• Every heading keeps its own line, with the paragraph break before and after it exactly as in the original. If a heading text reads like "1.1 Research Background", it must appear on its own line, not inside or at the tail of a sentence.
• Rewriting happens strictly INSIDE the body text between two headings; the headings themselves are protected elements (Gate B) and the boundaries are checked by Gate G.

(b) Merged mega-paragraphs. Do not combine paragraphs to fight cross-paragraph repetition. Merging produces over-long blocks (e.g. 212 words / 18 sentences) that are exhausting to read.

• Keep the original paragraph breaks. Cross-paragraph repetition (D13) is reduced by DIFFERENTIATING neighbouring paragraphs (changing each one's entry point, order, and function), NOT by merging them.
• A rewritten paragraph must stay within readable limits: about 160 words and about 9 sentences maximum (soft target), and must never exceed 200 words or 12 sentences (hard cap). If staying under the cap conflicts with reducing AI risk, keep the paragraph split — never merge to a longer block.
• "Fragment cluster" and "paragraph group" are DIAGNOSIS units (for planning cross-paragraph differentiation). They are NOT licence to output one merged paragraph. Each original paragraph remains a separate paragraph in the output.
• If a single original paragraph is itself longer than the hard cap, you may split it into two paragraphs at a natural logical break (this improves readability and adds no AI risk), but you may never merge.

Gate G (Phase 7) enforces both (a) and (b). Any output that violates them fails and must be redone.

Principle 7 — "Reorganisation" means within-paragraph only (never reorder paragraphs or change the argument)

Throughout this skill, "reorganise", "reorder", "change entry point", "paragraph architecture", and "functional reorganisation" mean changing the order and framing of sentences INSIDE a single paragraph, plus differentiating neighbouring paragraphs from each other. They do NOT mean:

• shuffling the sequence of paragraphs within a section;
• changing the logical development or argument structure of the chapter;
• moving content from one section to another.

The document's paragraph order and argument flow are preserved. The skill has no technique that rearranges whole paragraphs or alters how the argument unfolds, and it must never invent one — doing so would break logical coherence and the reader's ability to follow the thesis.

What IS allowed (reporting-order variation within a unit): inside a single paragraph you may change the entry point and the order of its sentences — for example, open with the problem or the result before the background, instead of the AI-typical "background → purpose → method → result → conclusion" order. This is encouraged (T18, T14) and is how "reordering the argument's presentation" is achieved safely. What is NOT allowed: moving paragraphs around, changing which section makes which point, or altering the chapter's overall line of argument. Reorder the presentation inside a paragraph; never reorder the paragraphs or rewire the chapter.

Principle 8 — Safe sentence splitting (no fragments, no choppy prose)

Splitting long sentences is a useful rhythm tool, but it must never produce grammatically incomplete fragments or a stop-start, half-sentence reading experience. Three failure modes are forbidden.

Detection method — UNIVERSAL completeness test. A fragment may start with ANY word and be ANY length. Do not detect fragments by matching a fixed list of opening words ("Including / Drawing on / Using / Based on / Considering ..."), and do not look only at a short 3–7 word window. Those approaches are a known failure mode that misses real fragments such as "Work engagement, and creative inspiration.", "Workplace communities, and internal corporate channels.", "Promoting, and implementing new ideas.", and any long "Including ..." continuation.

Instead, for every unit ending in a period, ask:

• Does it have its OWN grammatical subject?
• Does it have its OWN finite (tensed) main verb — not just an -ing / -ed participle, not just a to-infinitive?
• Does it express a complete thought on its own?

If any answer is no, it is a fragment, regardless of its first word or its length.

Common types (illustrative, NOT a closed match-list):

• Subordinate-clause fragments: "When employees interact with AI in daily work." — must attach to a main clause.
• List/enumeration tails: "...the mediating roles of self-efficacy. Work engagement, and creative inspiration." / "...through online social platforms. Workplace communities, and internal corporate channels." / "...dependent variable. Gender, age...".
• Subject-less verb phrases: "...it raises the problem. Builds the theoretical model.".
• Gerund/participle continuations: "...statistical analysis. Including multivariate analysis, ..." / "...the model. Using SEM to test the paths." — -ing with no finite verb.
• Choppy mechanical chopping into many ~10-word units that creates a robotic stop-start rhythm.

Rules:

• A sentence may be split ONLY when both resulting parts independently pass the subject + finite-verb test. The safe split test: cover everything before the proposed period and read what remains; if it lacks its own subject or its own finite verb, do NOT split there.
• NEVER place a period inside a list, before "and X", before a bare verb phrase, before an "Including / Such as / For example ..." phrase, or before any -ing/-ed continuation. In every such case keep it as one sentence.
• After splitting, each piece must stand on its own and read aloud as a complete sentence.
• Prefer a natural mix of short, medium, and long sentences. As a guide, keep most sentences roughly 8–30 words; an occasional shorter sentence is fine for emphasis, but avoid three or more ultra-short sentences in a row.
• Upper bound, too: Do not over-correct into long, tangled sentences. Keep sentences at or under about 30 words and at most two subordinate clauses. A sentence that grows long or heavily nested during rewriting must be split into two complete sentences.
• If reducing rhythm uniformity would require creating a fragment, do NOT split — vary the rhythm another way (reorder clauses within the sentence, change the opening, recast voice) instead.
• Gate H (Phase 7) applies the universal completeness test to every period and is output-blocking.

Principle 9 — Grammatical correctness and no jarring meta-commentary

Reducing the AI score must never introduce grammar errors or out-of-place asides. Two failure modes are forbidden:

(a) Grammar errors. The output must be correct academic English. In particular:

• Subject–verb agreement must hold (a singular subject takes a singular verb; a plural subject takes a plural verb). Rewriting that changes a subject from plural to singular (or splits a sentence) must update the verb to match.
• Tense, articles, prepositions, and pluralisation must remain correct.
• AI features must NEVER be reduced by deliberately inserting grammar mistakes or "non-native errors". Slight rhythmic plainness is fine; actual errors are not.
• Gate I (Phase 7) checks grammar, with special attention to subject–verb agreement at every point where a sentence was split, recast, or had its subject changed.

(b) Jarring meta-commentary. Do not insert sentences that explain, justify, or comment on the writing or the research design from outside the argument, such as:

• "This detail was included to improve measurement precision."
• "This sentence is added to clarify the logic."
• "The following paragraph explains the mechanism." These read as abrupt logical jumps and look machine-inserted. Rules:
• Every sentence must follow from the surrounding content and advance the thesis's own argument.
• Researcher-trace sentences (T19) are permitted only when they express the author's genuine reasoning within the local argument and connect smoothly to the sentences before and after. They must never be generic justifications of methodology, measurement, or of the rewrite itself.
• A researcher-trace or transition sentence that does not connect logically to both its neighbours is a defect — remove it or rewrite it so the paragraph reads as one continuous line of reasoning.
• Gate J (Phase 7) checks for inserted meta-commentary and logic jumps.

Chapter Version Ledger (required before every round)

Before starting any rewrite round, output and maintain this ledger. If the user cannot supply the previous ledger on a resumed task, rebuild it from the latest detector report before rewriting anything.

| Chapter | Current measured score | Lowest-ever score | Source version of lowest | Current escalation rung | Tried techniques on this rung | Rounds done | Below target (<30%)? | This round: rewrite? |
|---------|------------------------|-------------------|--------------------------|-------------------------|-------------------------------|-------------|----------------------|----------------------|
| ...     | ...                    | ...               | ...                      | 1–6                     | e.g. T17, T21                 | int         | Yes / No             | Yes (target) / No    |

Rules:

• The "rewrite?" column should normally say Yes for exactly ONE chapter per round: the highest current measured chapter that is still at or above the target (default 30%). It may say Yes for two chapters only if the user explicitly requests a two-chapter batch and agrees to re-test each chapter separately.
• Any chapter already below the target (default 30%) is marked done and is not rewritten further.
• After a re-test, update the ledger: record the new score, update lowest-ever, update the escalation rung, and mark "Below target?" Yes/No. If a chapter regressed, restore the lowest-ever source and record the rollback.
• The task is complete only when every chapter shows "Below target? = Yes". As long as any chapter is still at or above target, name the next round's target only as a post-retest candidate and do not present the work as finished; do not plan detailed rewrite actions for that candidate until the user provides the re-test result for the current round.
• Never start a new round without an up-to-date ledger.

This skill is only for reducing AI-writing-risk patterns and AI detector scores.

It is not designed for plagiarism reduction, similarity-score reduction, citation rewriting, or Turnitin Similarity Index optimization.

If the user asks for similarity reduction or plagiarism reduction, treat it as a separate task and do not mix it with AI-rate reduction.

Task Routing Rule

Before any diagnosis or rewriting, classify the user request and route it:

1. Skill-Audit Mode — the user asks to inspect THIS skill or its reference files for errors, contradictions, omissions, or optimisation. Skill Audit Mode is fully separate from Thesis Rewrite Mode: when the input files are SKILL.md or the reference modules, do NOT score them as thesis paragraphs and do NOT apply D1–D17 rewriting; audit them for internal consistency, routing completeness, safety boundaries, contradictions, missing execution gates, stale terminology, broken cross-references, and agent usability. Do NOT run thesis D1–D17 scoring and do NOT rewrite thesis text. Output rule-level problems, severity, affected files, suggested edits, and reasons, using this format:

   Issue ID:
   File / Section:
   Problem type: contradiction / omission / ambiguity / execution burden / unsafe wording / terminology drift
   Original wording:
   Recommended replacement:
   Reason:
   Priority: High / Medium / Low

   When the user asks for “原文 + 修改后内容” (original + revised), a compact variant is acceptable: Issue ID / File-Section / Problem type / Original wording / Recommended replacement, adding Reason only where needed for clarity. If the user provides a custom audit loop or table format, follow the user's format while preserving the core fields (location, original wording, problem, suggested fix, reason/priority). Skill-Audit Mode still must not run thesis D1–D17 scoring or rewrite thesis text.

2. AI-Writing-Risk Mode — the user asks to reduce AI-writing risk / AI rate / AIGC traces / Turnitin-AI / CNKI-AIGC / GPTZero, etc. This is the main thesis workflow (Diagnosis-Only / Working / Quick / Single-Batch / Full Staged below).

3. Similarity / Plagiarism Mode (out of scope) — if the user asks for similarity reduction, plagiarism reduction, Turnitin Similarity Index, 查重, 相似度, or 降重 with NO AI-writing context, do not proceed under this skill. Ask whether they actually want AI-writing-risk reduction instead. This skill never optimises text-similarity / 查重 scores.

4. Assembly Mode — only after chapter-by-chapter accepted versions already exist and have been externally re-tested. Assembly means combining locked accepted chapter versions into a report or DOCX. It does NOT mean rewriting a whole thesis in one pass. The capability assemble_locked_rewritten_document refers ONLY to this assembly of already-accepted chapters (and single-chapter/batch output for the current round), never to one-pass whole-thesis rewriting.

Language Routing Rule

• English thesis text: use D1–D17, the English cliché banks (Gate J), and the rewrite techniques.
• Chinese thesis text: use the Chinese-text AI-risk module (references/chinese_text_ai_risk.md) plus the structure-level checks; do NOT apply the Chinese-student-English cliché rules (those are for English text written by Chinese authors, not for Chinese text).
• Mixed Chinese–English thesis (e.g. English body + Chinese abstract/questionnaire): process by language segment, and protect bilingual term consistency (a term's Chinese and English forms must stay aligned).
• Method routing: apply the qualitative module (references/qualitative_authorship_restoration.md) and T25–T28 only to qualitative/mixed passages; do NOT apply qualitative researcher-voice techniques to a quantitative passage.
• Quotation-evidence protection (D21 and its Chinese-route equivalent): when a uniform-citation pattern is flagged, the default action is report-first. Before any rewrite planning, do NOT automatically change quotation count, quotation length, participant distribution, or theme coverage; only surrounding framing may be varied, and only where the existing analysis supports it (see detection_principles.md D21).
• Mixed-language paragraph: if a single paragraph contains both English and Chinese thesis prose, split it by language span where possible and score each span under its own route; protect bilingual technical terms, citations, and translated construct names before scoring, and never translate one span into the other language to unify the route.

For Chinese thesis text, the visible evidence must use the Chinese-route fields below (not a full English D1–D17 table):

Paragraph ID:
Language Route: Chinese
Chinese AI-risk groups hit: template opening / template closing / unsupported intensifier / repeated phrase block / copula avoidance / causal-chain stacking / inflated importance / empty comparison / high-frequency AI vocabulary / term-abbreviation repetition / Chinese qualitative uniformity (if applicable)
Shared structure risks: paragraph architecture / section-template dependency / reporting-sequence regularity / external-highlight fragment
Protected elements:
Risk estimate: chinese_paragraph_risk_estimate (internal Chinese-text risk estimate only)
Rewrite action: minimal edit / no rewrite / author confirmation needed

Task Mode Definitions

Mode routing summary (definitions below remain canonical; this table only consolidates them):

Before execution, identify the task mode.

Diagnosis-Only Mode

Use Diagnosis-Only Mode when the user asks to inspect, review, audit, identify risks, or list problems in thesis text without rewriting. If the user asks to inspect this skill, its workflow, or its reference files, use Skill-Audit Mode instead and do not run thesis D1–D17 scoring.

Rules:

• Do not rewrite thesis text.
• Output risk locations, reasons, severity, and suggested techniques only.
• Do not create a rewritten version unless the user explicitly asks for rewriting.

Working Mode

Default mode for normal thesis revision.

Rules:

• Show full execution evidence for rewritten high-risk (target) paragraphs. For each rewritten target paragraph, show the minimum evidence packet: Paragraph ID; route; hit dimensions or Chinese-route risk groups; selected techniques; protected elements before/after; exact Pass 1–3 actions; Gate A–J summary; original/revised word count; candidate status pending external re-test.
• For low-risk or skipped paragraphs, give a compact skip record — paragraph ID, section type, final risk estimate, primary skip reason, and protected-element status — not a full D1–D17 table.
• Include protected-element result and quality-gate summary.
• Do not overload the user with full D1–D17 evidence for every paragraph unless necessary.

Thesis-Evidence Audit Mode

Use this mode only when the user explicitly asks for full evidence, full audit, complete paragraph-level diagnosis, or detailed execution proof for THESIS TEXT. If the user asks to audit the skill file, workflow, or reference modules themselves, use Skill-Audit Mode instead. This mode controls evidence depth, not task size: a long task in this mode still follows Full Staged execution.

Rules:

• Show D1–D17 scoring.
• Show dimension-to-technique mapping.
• Show three-pass evidence.
• Show protected elements before and after rewriting.
• Show Gate A–J results for every rewritten paragraph or rewritten unit. Gates F–J are mandatory whenever text is rewritten, including quick single-paragraph mode.

Quick Mode

Use Quick Mode when EITHER (a) the user explicitly asks for a quick rewrite/check or says “just this paragraph”, OR (b) the user provides only one paragraph, no uploaded thesis/chapter/DOCX/PDF detector report or external-score report, and no external detector score. An optional same-author style sample, the single source paragraph file, or one short paragraph-level detector excerpt does NOT disqualify Quick Mode, but the uploaded file role must be stated explicitly. A thesis chapter, full thesis, or external detector report DOES disqualify it.

When Quick Mode is inferred rather than explicitly requested (case b), state: “Scope = Quick single-paragraph mode; no chapter-level or document-level score is produced,” so the user can redirect.

Quick Mode Anti-Misrouting Rule: do NOT use Quick Mode when the request mentions "chapter", "section", "thesis", "paper", "DOCX", "external report", "CNKI", "Turnitin", "full text", "整章", "全文", "论文", "检测报告", or includes any uploaded long document — even if the user also says "quickly" / "简单看一下". Route those to Full Staged Mode or Single-Batch Mode by paragraph count and evidence availability.

Quick Mode still requires minimum evidence: scope statement; hit dimensions (or Chinese-route risk groups); selected techniques; protected-elements check; a one-line Gate A–J summary for the rewritten paragraph; and the statement that no chapter-level or document-level score is produced.

Score Terminology Glossary

Use the following terms consistently.

• external_detector_score: the score reported by CNKI, Turnitin, GPTZero, Originality.ai, or another external detector.
• chapter_red_line: the lowest current measured external score for a chapter.
• raw_paragraph_risk: the internal D1–D17 weighted paragraph score before floors.
• paragraph_risk_index: the internal paragraph risk score after floors.
• overall_risk_index: the internal document-level risk estimate, not an external detector result.
• AI feature value: an external detector’s own reported value; record it as external evidence only.
• internal risk estimate: any score generated by this Skill without an external detector re-test.

Mandatory rule: Internal scores help locate risk. They must not be presented as proof that an external detector score has decreased.

Intent Confirmation Rule

Before starting Phase 1, check whether the user is asking for AI-writing-risk reduction or similarity/plagiarism reduction.

If the user mentions “Similarity Index,” “similarity,” “plagiarism,” “查重,” “相似度,” or “Turnitin Similarity,” respond with:

This skill only reduces AI-writing-risk patterns. It does not optimize plagiarism, similarity score, or Turnitin Similarity Index. Please confirm whether you want to continue with AI-writing-risk reduction.

If the user provides an AI detector report, AIGC report, Turnitin AI report, CNKI AIGC report, GPTZero report, or Originality.ai report, continue with AI-writing-risk reduction.

Triggers

• "降AI" / "降低AI率" / "降低AIGC检测率"
• "AI detection" / "reduce AI score" / "lower AI rate"
• "改写降AI" / "去AI痕迹" / "AI痕迹检测"
• Uploading an English academic paper with AIGC reduction intent
• User reports a mismatch between internal AI score and Turnitin/GPTZero/Originality.ai score

These triggers route the task to academic-integrity-preserving risk-pattern revision — not detector evasion, fabricated humanisation, or plagiarism/similarity manipulation. The skill reduces shared AI-writing risk patterns by minimal edits to the author's own text; it never bypasses detectors, fabricates content, or routes text through external rewriting services.

Universal AI-Feature Reduction Rule

This skill aims to reduce AI-writing features that are commonly sensitive across different AI detection systems.

It does not optimize for one detector. External detector reports are used only as evidence for locating risk areas.

The core reduction targets are:

1. repeated academic templates;
2. repeated paragraph function paths;
3. mechanical methodology and empirical reporting;
4. predictable reporting sequences;
5. uniform sentence rhythm and low burstiness;
6. over-smooth reasoning chains;
7. continuous AI-like fragments in abstract/introduction;
8. repeated hypothesis-development routes;
9. repeated result → implication loops in discussion or conclusion;
10. lack of human research judgment, boundary, decision trace, or process narration.

Mandatory rule:

• Convert detector-specific findings into shared risk categories before rewriting.
• Select rewrite techniques according to risk category, not detector name.
• Use section-level rewriting when the risk appears across paragraph groups.
• Do not rely on synonym replacement, random short sentences, or general polishing.
• Preserve all protected elements, including citations, variables, hypotheses, coefficients, p-values, model names, table numbers, and statistical conclusions.

Validated Researcher-Trace Rule

Validated finding:

• In prior task testing, researcher-view judgment sentences and process narration appeared useful for some descriptive, overly smooth, abstract, or mechanically explanatory paragraphs. Treat this as a conditional strategy, not a universal fix.

Mandatory rule:

• This strategy must be treated as a formal Phase 4 rewrite strategy.
• T19 = source-extracted relocation ONLY: it may move, split, or lightly rephrase researcher-side judgment/process/boundary material that ALREADY EXISTS in the thesis. If a sentence is newly COMPOSED from existing thesis facts rather than relocated, classify it as T29 (evidence-based reconstruction), not T19. Either way a completed Source Trace block is required; if no source can be identified, do not write the sentence.
• It should be marked as a High-Efficiency Strategy.
• Use this strategy only when the added sentence can be directly grounded in existing thesis content. Do not insert researcher-trace sentences into statistical-result paragraphs unless the sentence is tied to the reported table, coefficient, sample, model, measurement choice, or stated finding.
• The added sentence must be based on existing thesis content. It must not introduce new data, new theory, new references, or unsupported examples.

Project-Level Working Structure

For long thesis documents, use a two-level working structure:

1. Master overview

   • thesis title, chapter structure, total paragraph count
   • current external AI detector result if available, including detector name, score type, score value, and highlighted section evidence
   • document-level AI risk summary
   • high-risk sections and rewrite priority
   • chapter status matrix

2. Chapter task notes

   • section type
   • detected AI patterns
   • paragraph-level risk points
   • rewrite strategy
   • completion status

This structure is especially useful when the thesis has more than 50 pages or when any external AI detector score is above 30%.

Mandatory project-state rule:

• If the task involves multiple chapters, update the Master Overview after each completed round's Stage 4 final output and before the Stage 5 mandatory re-test stop. Do not use "Stage 4" to mean "Phase 4": Stage 4 is the round-level final-output stage, while Phase 4 is target selection / prioritisation. Execution order is locked — the Phase workflow runs inside the round first, then the Stage output. A task may not jump from Phase 7 quality gates directly to Done; it must pass Stage 4 scope-aware output and then the Stage 5 external re-test stop, unless the user explicitly selected Quick Mode with no detector-score objective.

Terminology Rule: Phase denotes the executable workflow steps and is the ONLY control of execution order; Stage is a user-facing summary grouping and must never override, skip, or reorder Phase execution. Mapping: user-facing Stage 1 = Phase 1–3; Stage 2 = Phase 4–5.5; Stage 3 = Phase 7; Stage 4 = Phase 8–9; Stage 5 = the external re-test checkpoint.

• The Master Overview must include chapter status, completed batches, remaining high-risk sections, baseline version, latest external score, and next recommended batch.
• At the start of a resumed task, use the Master Overview as the recovery context.

Global D1–D17 Consistency Override Rule

This skill uses a formal D1–D17 core scoring system for general and quantitative AI-trace risk, plus the qualitative-only supplement D18–D21 for qualitative and mixed-methods papers.

All references to “15 dimensions,” “D1–D15 coverage,” or “42.0 total score” must be treated as outdated unless explicitly marked as historical notes.

Current scoring system:

• For English-route thesis text, D1–D17 must be evaluated for every included body paragraph in the current target chapter or selected rewrite batch. For Chinese-route thesis text, do not force the English D1–D17 table; use the Chinese-route evidence template plus shared structure-level checks. For mixed-language documents, apply the required route separately by language segment. "Included body paragraph" does not mean every paragraph in the whole thesis unless the user explicitly requests full-document diagnosis.
• Total possible dimension score = 49.0.
• raw_paragraph_risk = min(dimension_score / 49.0, 1.0) × 100%.
• Phase 2 is incomplete if D16 or D17 is missing.
• Public evidence is incomplete if the Detection Report does not show the required route-specific scoring evidence: English-route paragraphs require D1–D17 scores; Chinese-route paragraphs require the Chinese-route evidence template plus shared structure checks; qualitative or mixed-methods passages require the applicable D18–D21 supplement or its Chinese-route adaptation.
• D12, D13, D14, D15, D16, and D17 must always be checked manually or semi-manually.

If any rule conflicts with this 17-dimension system, the 17-dimension rule overrides the older 15-dimension wording.

Score Regression Guard

Before rewriting, identify the current lowest external AI-score version as the baseline version, using the Chapter Version Ledger (see Core Operating Architecture).

If multiple rewritten versions exist, do not assume the newest version is better. Use the version with the lower external detector score as the working baseline. This applies per chapter, not only to the whole document — keep the lowest-scoring version of EACH chapter.

For long thesis documents, do not rewrite the whole document before testing a small batch. Rewrite one chapter, re-test it, then continue (Principle 1).

Recommended process:

1. Build/refresh the Chapter Version Ledger and identify each chapter's lowest-ever version.
2. Target-lock the highest current measured chapter that is still at or above the target (default 30%); skip any chapter already below target.
3. Diagnose their D1–D17 risk dimensions.
4. Rewrite only those chapters by minimal edit (never regeneration).
5. Run protected elements check.
6. Run Phase 7 quality gate.
7. STOP and have the user re-test the rewritten chapter with the external detector.
8. After a comparable external re-test, handle the chapter by status:
   • If the score is below the target threshold: mark the chapter DONE and lock that version within the named detector stream and scope.
   • If the score decreased but remains at or above the target: record it as PROGRESS, set it as the new lowest-score baseline for that chapter, and plan the next escalation rung for the same chapter. A decrease that is still above target is progress, not completion.
   • If the score stayed the same or increased: mark the candidate rejected, roll back to the chapter's lowest-ever version, diagnose the failed rewrite direction, and redesign the next small-batch strategy.

Mandatory rule:

• A chapter's current measured external score is a RED LINE. A rewrite must be treated as provisional until the user provides an external re-test result. If the re-test does not measurably lower the score, discard the rewrite and restore the chapter’s lowest-ever external-score version.
• If Version A has a lower external AI score than Version B, continue from Version A unless there is a clear academic-quality reason not to.
• Do not continue rewriting from a version that increased the external AI score.
• If a sample rewrite increases AI risk, stop and redesign the rewrite strategy.
• If the external AI score increases after a rewrite, return to the lower-score baseline version of that chapter.
• Score increase triggers the same action path: roll back to the chapter's lowest-ever version → diagnose failed rewrite direction → redesign strategy → retest with a small batch.
• Stalemate vs Rewrite Fatigue (two separate thresholds; Rewrite Fatigue Pause overrides Stalemate Warning): Stalemate Warning fires at <5 percentage points change across two consecutive comparable re-test rounds — mark the strategy stale and plan the next rung, but do NOT auto-pause. Rewrite Fatigue Pause fires at <3 percentage points change across two consecutive rounds, or once three complete rewrite+re-test rounds are done on the chapter — mark the unit Paused unless a clearly untried lower-risk technique remains. Stalemate handling: if a chapter changes by less than 5 percentage points across two consecutive comparable re-test rounds, mark the current strategy as stale, keep the lowest-score version, and plan the next round at the next escalation rung. If the next rung stays intra-paragraph and preserves all structural boundaries, user confirmation is not required beyond the normal external re-test checkpoint. If the next rung would require broader structural reorganisation beyond the current approved unit, ask for user confirmation before execution.

Reference Availability Check: before execution, confirm the four reference modules (references/detection_principles.md, references/rewrite_methods.md, references/qualitative_authorship_restoration.md, references/chinese_text_ai_risk.md) are readable. If any module is missing or unreadable, mark the run "Partial Execution — reference module unavailable", name the missing module, do not produce final scoring that depends on it, and continue only with the parts that remain fully supported.

Reference Path Rule: in packaged skill distribution, load reference files from references/. If the skill is instead being audited from uploaded standalone files, match by canonical basename first (detection_principles, rewrite_methods, qualitative_authorship_restoration, chinese_text_ai_risk) and ignore upload suffixes such as (15) or copy markers.

Score Terminology Glossary: external_detector_score = the score reported by CNKI, Turnitin, GPTZero, Originality.ai, or another external detector. chapter_red_line = the lowest current measured external score for a chapter. raw_paragraph_risk = the internal D1–D17 weighted score before floors. paragraph_risk_index = the internal paragraph risk score after floors. scope_risk_index = the internal risk estimate computed over whatever scope was diagnosed; it is reported as overall_risk_index only for a fully diagnosed full-document scope (otherwise as target_chapter_risk_index / batch_risk_index / fragment_risk_index), and is never an external detector result. AI feature value = an external detector’s own reported value and should be recorded as external evidence only. Priority notation rule: use either A/B/C/D or P0/P1/P2/P3 within one report, not both; if legacy notation appears, the mapping is exactly A=P0, B=P1, C=P2, D=P3.

Target Status State Machine

Every chapter/target-unit's status at the end of a round is reported as exactly ONE of these six labels: Done / Progress / Rejected / Paused / Blocked / Incomplete. Rejected means a comparable re-test regressed or failed to improve against the red line, so the candidate is discarded and the lowest-ever version remains the baseline. Partial Execution is NOT a seventh target-unit status: it is a run-level execution label used only when one or more required reference modules are unavailable. When it applies, the affected target-unit status is Incomplete (unless the missing module is irrelevant to the selected route), and the run must name the missing module and which scoring/rewriting functions were skipped. This is a reporting layer over the existing stop-logic (Principle 0 pause reasons, Rewrite Fatigue Detection, Score Regression Guard); it changes none of those rules, it only standardises how the outcome is named so a unit is never ambiguously "stuck".

• Done — the unit's latest COMPARABLE external re-test is below target (default <30%). Only an externally-confirmed below-target unit may be Done. An internal estimate alone can never mark Done.
• Progress — the latest comparable re-test decreased versus the chapter red line but is still ≥ target. Continue next round (move up the escalation ladder if <3pp gain).
• Rejected — a comparable re-test regressed (higher than the red line) or did not improve against it. Discard the candidate, keep the lowest-ever comparable version as the baseline, diagnose the failed rewrite direction, and escalate one rung on the next round after recording the rollback. A regression is never reported as Progress.
• Paused — work is intentionally halted for a legitima

Truncated - read the full file at https://github.com/Moonlit-Pages/AIGC-Detector-Rewriter-Skill/blob/84f93cd4ac15477e363229e0dfa690091bb689d9/SKILL.md.