human-writing

这段内容是审核时保存的快照。外部来源才是完整且最新的版本。

---
name: human-writing
description: >
  Write text that reads as authentically human rather than AI-generated. Use this skill for ANY
  writing task where the output must not sound like it came from an AI: blog posts, emails, essays,
  articles, creative writing, marketing copy, social media posts, reports, documentation, cover letters,
  bios, product descriptions, or any content that will be read by humans who might notice or care
  about AI tells. Also trigger when the user says things like "make it sound human", "write naturally",
  "don't sound like AI", "avoid AI slop", "humanize this", or "write like a real person". This skill
  should be used even when the user just asks for writing and doesn't mention AI at all, if the context
  suggests the output will be published, shared, or evaluated by humans.
---

# Human Writing Skill

Write like a person, not a language model. This skill is built from peer-reviewed research
(UCC stylometry study, Dec 2025; Carnegie Mellon detection analysis, 2025; Kujur comparative
linguistics study, SSRN Nov 2025), Wikipedia's "Signs of AI Writing" editorial guide, the
Max Planck Institute vocabulary frequency analysis, Buffer's 52M-post corpus analysis, and
the 2026 arxiv paper "The Last Fingerprint: How Markdown Training Shapes LLM Prose."

The core finding across all of this research is that AI detection operates on three axes:
**perplexity** (word-level predictability), **burstiness** (sentence-level variation), and
**stylometric fingerprinting** (vocabulary and structural patterns). Human text scores higher
on the first two and is far more varied on the third. This skill addresses all three.

---

## HOW TO USE THIS SKILL

Before writing anything, read this file completely. Then read `references/banned-vocabulary.md`
for the full word/phrase avoidance list. Apply every rule below to all output. After drafting,
do a self-audit pass (Section 7) before delivering to the user.

---

## 1. VOCABULARY RULES

AI models converge on a small rotation of "impressive-sounding" words that now function as
statistical fingerprints. The Max Planck Institute found that words like "delve," "underscore,"
and "pivotal" spiked 50%+ in published text after ChatGPT launched. Wikipedia editors maintain
a living list of AI vocabulary grouped by model era.

### 1.1 Banned words (never use these)

The full list is in `references/banned-vocabulary.md`. The worst offenders, which must never
appear in output regardless of context:

delve, tapestry, testament, landscape (figurative), vibrant, pivotal, multifaceted, nuanced
(as filler), robust, seamless, cutting-edge, groundbreaking, game-changer, paradigm shift,
holistic, synergy, leverage (verb), foster, facilitate, utilize (use "use"), underscore,
spearhead, embark, unpack, unravel, craft (verb, when meaning "write" or "make"),
comprehensive, invaluable, endeavor, realm, intricate, meticulous, plethora, myriad (noun),
paramount, resonate, streamline, elevate, empower, unlock, unleash, harness

### 1.2 Banned phrases (never use these)

"In today's [adjective] [landscape/world/era]", "It's worth noting that", "It's important
to note", "Let's dive in", "Let's delve into", "At the end of the day", "Not just X, but Y"
(the false-contrast pattern), "In the realm of", "A testament to", "Serves as a reminder",
"It cannot be overstated", "This is where X comes in", "When it comes to", "The question
is not whether but how", "As we navigate", "In an era of", "The landscape of", "Moving
forward", "Going forward", "On the journey to", "In the ever-evolving world of",
"Here's the thing", "The reality is", "The truth is", "Make no mistake"

### 1.3 Banned openers (never start a sentence or paragraph with these)

Certainly, Moreover, Furthermore, Additionally, Indeed, Notably, Importantly, Interestingly,
Remarkably, Ultimately, Fundamentally, Essentially, Undoubtedly, Consequently, Subsequently,
Accordingly

### 1.4 What to use instead

Use plain, specific language. Prefer short A