Why use it: the biggest risk of the Humanizer is not under-editing, but over-editing — stripping away the real voice, removing a legal disclaimer, or rewriting a quotation. SKILL.md already has two guardrails for this; this file shows how to invoke them directly.
What it produces: a narrower-scoped run that keeps what is worth keeping.
The commands in this file are typed in the agent's chat window.
The "must not be flagged" list
SKILL.md → "Check for false positives" lists things that are not in themselves evidence of AI. Most memorable:
- Standard grammar, consistent style. Many writers are professionals or have been through editing.
- Formal, academic words. §7 targets only a specific word list, not every formal word.
- A curly quote on its own. macOS, Word, Google Docs curl quotes by default.
- An em dash on its own. Many editors and journalists use it frequently. It only counts as a signal when paired with clichéd advertising-style rhythm.
- A short sentence for emphasis. Only flag when several dramatic sentence fragments are strung together.
- Deliberate repeated paragraph openings. "She came. She saw. She conquered." is rhythm, not an error.
- "Honestly" or "look" mid-sentence. The signal is the dramatic opening, not the word itself.
- Useful limits and disclaimers. Keep scope statements, legal and safety warnings, genuine corrections, sourced counterarguments, FAQ answers.
- Real alternatives. In a design doc or tutorial, keep the options the reader might weigh. Only drop far-fetched alternatives that were dismissed and never mentioned again.
- Quoted text. Do not rewrite a tracked phrase when it is inside quotation marks, a title, a proper name, or an example being discussed.
- Revisions before 30/11/2022. The date ChatGPT was released to the public.
General rule: "When unsure, look for several patterns together. One em dash proves nothing."
Details people need to keep
SKILL.md → "Human details to keep" — the things that carry a person's voice, which must be kept unless they break the meaning: strange and specific details, mixed emotions not yet resolved, slang tied to particular years, deliberate first-person choices, uneven sentence lengths, and sentences that interrupt themselves in parentheses.
How to name that guardrail clearly
When the returned version sounds blander than the original:
Bản viết lại đã bào mất giọng. Chạy lại và bám mục "Human details to keep" —
giữ các asides, câu tự sửa mình, và nhịp câu dài ngắn không đều của bản gốc.
When it strips out something that must be kept:
Bạn đã bỏ đoạn giới hạn phạm vi ở cuối. Theo mục "Check for false positives",
disclaimer có ích phải giữ nguyên. Trả lại đoạn đó rồi đưa lại bản cuối.
When it rewrites a quotation:
Đoạn trong ngoặc kép là trích dẫn nguyên văn, thuộc mục "Secondhand text" —
không được sửa. Chỉ chữa phần văn xuôi quanh nó.
Lock the scope from the start
Cheaper to block up front than to fix afterward:
/humanizer
Văn kỹ thuật, giữ trung tính, không thêm cá tính.
Không đụng gì trong ngoặc kép và trong code block.
Chỉ sửa §7, §14, §23.
[văn bản]
Check by eye: is any claim missing
Step 3 in the process makes the skill ask itself whether it added or lost any facts. You should verify independently, especially in file mode:
git diff --word-diff -- docs/launch-post.md
Read separately the spans that were deleted. Every name, number, date, or citation that is lost is an error under the skill's own rules — report it and make it restore them.