Slopbeth：移除 AI 写作痕迹的文稿编辑技能（slopkit/skills/slopbeth，v1.4.1）

Load references/evaluation.md for the full benchmark and detector-evidence rules. In this package, use scripts/ relative to the installed Slopbeth skill directory. The scripts report signals. They do not decide whether prose is good enough. Hard rules

• Never claim text is permanently undetectable, guaranteed human, or safe against all AI detectors.
• Reject detector tricks that make the writing less true, less specific, or less like the author.
• Keep vague copy evidence-bound. If concreteness requires missing source material, ask for it or label the example as a placeholder.
• Do not launder vague outcomes into polished claims. If the source gives only abstract benefits, name the missing mechanism, owner, metric, changed step, or evidence instead of restating the benefit as true.
• Leave support, recruiting, incident, product, strategy, and education copy without invented owners; dates; failure modes; workflow steps; product surfaces; company names; metrics; or obligations.
• In support copy, do not add process promises such as "we will review," "we will follow up," or "we will resolve" unless the source says that team action is available. Ask for the required next input and preserve promise boundaries.
• In policy and incident copy, do not add quality labels such as "auditable," "secure," "resilient," or "controlled" unless the source states that property directly. Keep the rule or incident boundary concrete.
• Preserve qualifiers that carry scope; uncertainty; causality; risk; or legal/technical meaning.
• Avoid replacing AI slop with a new formula: clipped aphorisms; tidy triads; forced contrast; dramatic fragments; or generic consultant voice.
• Over-editing already strong human text is a failure. A light edit or "leave this alone" can be the correct output.
• Instructions about the writing are not the writing. If the source says what the piece should or should not do, do it; do not print it. "Leave this alone" never means "hand the brief back".
• Mark exact spans when reviewing long or risky text: bad span; label; reason; preserved span; reason. If the exact span cannot be pointed to, treat the critique as too vague.
• Check cadence before finalizing medium or long rewrites. Repeated sentence lengths, polished transition stacks, and repeated openers can be slop even when the words are not banned.
• Avoid em dashes, emojis, title-case hype headings, and decorative bold unless the user's sample clearly uses them and the medium calls for them.
• Keep the skill's internal checklist shape out of final prose. User-facing rewrites should not default to title-case sections; labeled vertical lists; exhaustive caveat blocks; or polished three-part scaffolds.
• For detector-facing work, record structured rows with tool name; URL; date; text hash; raw result or screenshot path; result class; and limitation.