name

de-ai-slop

description

Remove the tells that make text read as AI-generated ("slop"), in Chinese, English, or Japanese, while preserving the author's own meaning and style. Use when the user asks to de-AI / humanize / 去AI味 / remove AI flavor from a draft, when rewriting/polishing/translating text that reads like an LLM wrote it, or when reviewing a draft for AI tells. Style-agnostic subtraction only — it strips AI patterns, it does not impose a voice.

de-ai-slop

A style-agnostic patch tool: it removes the structural tells that make a reader think "a model wrote this," without flattening the text into neutral documentation. It does not prescribe a voice — pair it with a voice/persona skill if you want one.

What this is NOT

Not a style guide. It subtracts AI tells; it never adds a particular voice, classical flourish, or "intellectual" register.
Not a rewriter-from-scratch. Patch the hits, keep the author's argument, paragraph functions, and edge.

Workflow

Detect the language of the target text (zh / en / ja) and load the matching reference. Mixed-language → load both.
Scan, don't carpet-bomb. Catch the top hits by severity (below), not every theoretical match.
Minimal patch per hit. Delete the template/signpost/inflation; fold the information back into the prose. Never expand one fix into a full rewrite.
Re-read for over-correction. If paragraphs thinned, sentences got short and uniform, or the author's judgment flattened into a manual — roll back.

Task dispatch

Rewrite / polish a draft → load the language's *-patterns.md, apply the minimal-patch method.
Translate → load references/translation-guardrails.md (structure-faithful, no added headings, no editorializing).
Review / score → load references/quality-rubric.md; report the top 5–10 hits by severity with an evidence quote each, plus a 0–100 score.

Severity (clear CRITICAL first)

CRITICAL — em-dash overuse (en/ja), uniform sentence length (no burstiness), fabricated-scenario openings, copula avoidance clusters (en), ですます単調 (ja). These give it away instantly.
HIGH — significance inflation, hedging clusters, AI vocabulary, the 二分对照壳 (不是A而是B), 强制三分割 (ja), negative parallelisms.
MEDIUM / LOW — Title Case, curly quotes, emoji, boldface, filler.

Universal rule (all languages)

No fabricated scenarios. Never open with, or argue from, an invented anecdote ("one Monday I saw…", "imagine a user who…"). Open from real experience, checkable material, or a direct claim. (zh rule 25, en/ja mirror.)

References

references/zh-patterns.md — Chinese, 33 numbered patterns + minimal-patch method.
references/en-patterns.md — English (burstiness, em-dash, copula, bloat, hedging…).
references/ja-patterns.md — Japanese (翻訳調, ですます, 接続詞, クッション言葉…).
references/quality-rubric.md — severity + 0–100 scoring, evidence-quoted.
references/translation-guardrails.md — structure-faithful translation.
references/SOURCES.md — credits for the community skills and references merged here.

de-ai-slop

What this is NOT

Not a style guide. It subtracts AI tells; it never adds a particular voice, classical flourish, or "intellectual" register.

Not a rewriter-from-scratch. Patch the hits, keep the author's argument, paragraph functions, and edge.

Workflow

Detect the language of the target text (zh / en / ja) and load the matching reference. Mixed-language → load both.

Scan, don't carpet-bomb. Catch the top hits by severity (below), not every theoretical match.

Minimal patch per hit. Delete the template/signpost/inflation; fold the information back into the prose. Never expand one fix into a full rewrite.

Re-read for over-correction. If paragraphs thinned, sentences got short and uniform, or the author's judgment flattened into a manual — roll back.

Task dispatch

Rewrite / polish a draft → load the language's *-patterns.md, apply the minimal-patch method.

Translate → load references/translation-guardrails.md (structure-faithful, no added headings, no editorializing).

Review / score → load references/quality-rubric.md; report the top 5–10 hits by severity with an evidence quote each, plus a 0–100 score.

Severity (clear CRITICAL first)

CRITICAL — em-dash overuse (en/ja), uniform sentence length (no burstiness), fabricated-scenario openings, copula avoidance clusters (en), ですます単調 (ja). These give it away instantly.

HIGH — significance inflation, hedging clusters, AI vocabulary, the 二分对照壳 (不是A而是B), 强制三分割 (ja), negative parallelisms.

MEDIUM / LOW — Title Case, curly quotes, emoji, boldface, filler.

Universal rule (all languages)

References

references/zh-patterns.md — Chinese, 33 numbered patterns + minimal-patch method.

references/en-patterns.md — English (burstiness, em-dash, copula, bloat, hedging…).

references/ja-patterns.md — Japanese (翻訳調, ですます, 接続詞, クッション言葉…).

references/quality-rubric.md — severity + 0–100 scoring, evidence-quoted.

references/translation-guardrails.md — structure-faithful translation.

references/SOURCES.md — credits for the community skills and references merged here.