菜单
Rewrite AI-generated text into genuinely human prose with seven ordered transform passes: add real named specifics, fix the structure, kill the tells (em-dashes, intensifiers, negative parallelism, meta-signposts), admit real limits, and read it aloud.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/ariaxhan/the-agent-library --skill humanize-ai-prose命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
Turn a borrowed metaphor, analogy, or cross-domain inspiration into a testable mechanism, or prove it is decoration and discard it. Use when a design leans on language from another field ("immune system", "evolution", "market", "quorum", "crystallization"), when someone proposes copying how another domain solves a problem, or when reviewing a system whose names sound deeper than its code. Produces either a named domain-free invariant with evidence and falsifiers, or an explicit rejection with the reason it failed.
Design systems so a failure can never present as success. Use when building or reviewing pipelines, extractors, reconciliations, checkers, or any computation whose output humans act on; when missing inputs, unreadable files, or crashes could pass silently; when a tool "found" an explanation that feels too convenient; or on "make this fail loudly", "could this lie to us", "harden the failure paths". Produces named invariants, explicit skip and no-answer states, honest cause classes, and adversarial regression fixtures.
A compact decision protocol for non-trivial choices, reviews, governance calls, and live disagreement. Use implicitly at decision points, even when nobody says "debate", whenever a choice is high-impact, hard to reverse, policy-bearing, or has multiple plausible paths that evidence alone cannot settle. Skip for trivial, cheap-to-reverse, or deterministically decidable choices.
基于 SOC 职业分类
| name | humanize-ai-prose |
| description | Rewrite AI-generated text into genuinely human prose with seven ordered transform passes: add real named specifics, fix the structure, kill the tells (em-dashes, intensifiers, negative parallelism, meta-signposts), admit real limits, and read it aloud. |
A reusable, ordered editing system that turns model-generated prose into writing that reads like an actual person wrote it. It's two layers: seven mechanical transform passes (each tell is detectable and rewritable), then a taste checklist of judgment calls no find-and-replace can make.
The passes run truth → structure → paragraph → sentence → word → friction → ear, and the order is load-bearing. Inserting one real specific reshapes everything downstream (it forces you to delete the vague sentence that was standing in for it), so specifics come first. Word-level fixes come late, because there's no point polishing a sentence you're about to cut. Read-aloud is last, after the text has stopped changing. Run them in order; don't jump to the word pass because it feels productive.
1. TRUTH: insert real specifics. AI prose names categories, never instances ("various campuses," "a meaningful experience," "a typical user"). Give every section at least one true, named, datable detail: a first name, a place, a month-and-year, a product, a number that isn't round. This is the highest-leverage move in the whole system. If you don't know the fact, do not invent it. Use the bracket rule below.
2. STRUCTURE: de-signpost, fix intro and ending. Delete sentences whose only job is pointing at other sentences ("Here's the part that surprised me," "In this article we will," "That's the whole flaw, stated cleanly"). Humans just say the thing. Rewrite the intro and conclusion completely rather than patching them: they're the most template-shaped parts of any draft. The ending should add something (an implication, an open question, a last concrete image), never re-list the points. Cap dramatic one-line closers ("mic drops") at two per piece; one is voice, five is a metronome.
3. PARAGRAPH: break the isomorphism. When every paragraph ends by extracting its lesson and tying it to the thesis, that uniform closure reads as machine even if each sentence is fine. Let at most a third of paragraphs end on a significance statement; let the rest end on a fact, an image, or nothing. Vary paragraph openers (scan the first word of each in a column: no grammatical shape twice in a row) and paragraph sizes (include at least one one-sentence paragraph and one 5+ sentence paragraph; open one with a question).
4. SENTENCE: restore burstiness in both directions. Human writing spikes in sentence length; AI flatlines, usually at 18–24 words, sometimes (post-editing) at a relentless punchy 6–12. Find any three consecutive sentences of similar length, cut one under 8 words, let another run 30–45 with subordinate clauses. Restore parenthetical asides (2–4 per page: they hold the spoken register flat prose lacks). Break decorative rule-of-three triads ("skill, compassion, and steadiness"): keep one item and make it concrete. Allow at most one negative parallelism ("it's not X, it's Y" / "that's not a glitch, it's...") per piece; rewrite the rest as direct statements that just say what it is.
5. WORD: lexicon and punctuation purge. Run this for human readers, not detectors (see below). Zero em-dashes: replace with a period, a comma, or parentheses, not a semicolon spree. Delete the abstract intensifier lexicon: profoundly, transformative, unwavering, invaluable, leverage, delve, harness, streamline, robust, seamless, nuanced, multifaceted, landscape, realm, tapestry, testament, crucial, pivotal, groundbreaking. Replace each with the plain word for the actual thing. Contract by default ("that's," "it's," "doesn't") and cap full-form "It is / That is / Here is" sentence-openers at two: surgical apostrophe-avoidance reads as machine. Delete cliché connectors (Furthermore, Moreover, Additionally, In conclusion, Ultimately); paragraph breaks do the connecting.
6. FRICTION: admit what was hard, boring, or wrong. AI prose is frictionless: nothing fails, drags, or stays unresolved. Readers flag that false omniscience faster than any word choice. Per piece, include at least: one thing that was boring or unglamorous, one limit or failure stated honestly, and one uncertainty left standing ("I'm not sure this scales past one case" beats fake confidence). Every admitted flaw must be true: invented friction is worse than none.
7. EAR: read it aloud. No transforms, just a final gate. Read it out loud (or subvocalize slowly). Any sentence you stumble on, or wouldn't say to a person across a table, gets rewritten in the words you'd actually use. Any sentence that exists only to sound good gets cut. Human looseness is allowed (a comma splice, a sentence a beat too long); typos and broken grammar are not.
As of 2026, the strong detectors (GPTZero, Pangram, Binoculars-style) score statistical structure: sentence-length variance ("burstiness"), paragraph shape, token predictability, transition-word frequency, not vocabulary. So the structural passes (2–4) move detection scores 20–35 points; the word pass (5) moves them under 5. If you're trying to pass a detector, the structure work is what matters. The word pass is for the human reading it. Do both, but know which job each one is doing.
The dangerous tells are the second-generation ones, because the obvious lexical ones (em-dashes, "delve," triads) are already widely scrubbed: uniform punchiness, uncontracted "It is / That is" anaphora, frictionless confidence, and missing specifics. The single highest-leverage fix remains one true, named detail per section.
After the passes, walk the judgment calls in reference/checklist.md. A "no" sends the piece back.
The three that matter most:
When a section needs a real specific (Pass 1) or an honest limit (Pass 6) and you don't have the fact, do not fabricate it. Insert a bracket placeholder that names exactly what's missing and hand it back to the author:
[BRACKET: need one real moment here: a name, a date, one thing that actually happened. A first name and one sentence beats a paragraph of reflection.]
The piece is not done while a bracket is open. This is the line between humanizing and lying: a humanized draft surfaces what only the author can supply; it never makes it up.
