skill — human-writing

A skill that makes LLM output read as naturally human by avoiding the tells cataloged in Wikipedia’s Signs of AI writing field guide (WP:AISIGNS) — the same page featured in the @yatesvids “Signs of AI writing” Instagram reel.

The field guide exists to help editors detect AI-generated text. This skill runs it in reverse: instead of detecting the tells, it stops the model from producing them — the AI-vocabulary words (“delve”, “vibrant”, “tapestry”, “pivotal”…), the significance inflation, the “not only X but also Y” constructions, the em-dash and boldface tics, the chatbot residue (“I hope this helps!”, “it’s important to note”, “[Your Name]”). The result is prose with the documented signs of human writing: plain “is/has” constructions, plain verbs, natural word repetition, and specific facts instead of filler analysis.

Contents

Install

Works with any agent that supports Anthropic-style skills (Claude Code, Claude.ai skills, Cursor, and other SKILL.md-based agents):

• Claude Code / repo skills: put this repo’s folder in ~/.claude/skills/human-writing/ (or copy SKILL.md into a skills/ folder of a project), then ask for anything and the model will apply it when relevant.
• Any other agent: point it at SKILL.md, or paste the contents as instructions.
• Plain prompt use: the SKILL.md body works standalone as a system prompt.

Usage

Trigger naturally with requests like:

• “write this essay so it doesn’t sound like ChatGPT”
• “make it sound human / remove the AI tone / humanize this”
• “rewrite my article and avoid the Wikipedia signs of AI writing”

The skill also includes a mandatory self-check pass: before delivering, the model scans its own draft for the tell list (banned vocabulary, “serves as”, negative parallelisms, tacked-on “-ing” analysis, emoji, spaced em dashes, placeholders…) and fixes any hits.

Testing

The repo ships a tell checker that encodes the skill’s rules as regexes, so the skill’s claims are verifiable:

# 1. List sync: the checker's banned vocabulary must match SKILL.md's table exactly
# 2. Detection: every banned entry and context rule must actually fire
python3 tests/check_tells.py --self-check --test-detection

# 3. "Before/after" demonstration:
#    ai_style.txt    -> should FAIL (it's the un-skilled, LLM-flavored draft)
#    human_style.txt -> should be CLEAN (the same content after applying the skill)
python3 tests/check_tells.py tests/samples/ai_style.txt tests/samples/human_style.txt

Expected: the self-check and detection test pass, ai_style.txt reports hard tells (banned words, “not only…but also”, emoji, spaced em dashes, “it’s important to note”, tacked-on “-ing” flourishes…), and human_style.txt reports CLEAN. The checker is a supporting tool, not a detector to trust blindly — like Wikipedia’s own caveats, no automated check is definitive; it exists to catch regressions in the skill’s examples.

License

MIT (see LICENSE). The reference catalog draws on Wikipedia:Signs of AI writing (CC BY-SA 4.0); attribution and links in reference/ai-writing-tells.md.