Human eyes
A Claude Code skill that audits prose for the habits readers now clock as machine-written
Language nerds are quick to defend writing techniques over-used by AI, such as parallelism, the rule of three, contrast framing, and (ahem) em dashes, because they also appear in good human writing.
The difference is intent. Human writers reach for an em dash to mark an interruption, an aside, a change of direction. AI models reproduce punctuation widespread in their training data.
The writing patterns are not the problem. They are symptoms of a lack of intent. As any Wittgenstein-pilled philosopher will tell you, text written without intent has no soul. As everyone else knows, it’s just irritating and vaguely unsettling to read.
I despise AI writing, but it’s here now. I would like to make it better.
What it does
Enter /human-eyes, a Claude skill for cleaning up AI-contaminated content.
Give it text and it audits against about 55 evidence-based patterns often bungled or over-used by AI. It points to the specific examples in your text so that you, or your agent, can write it better.
How it works
85% of the patterns are programmatically detectable via a script. The rest relate to overall structure, tone and content, and are assessed by your agent. That split reduces agent improvisation and token usage, and enforces a consistent approach and output across sessions and models.
I’m neurodivergent with a Claude 20x Max plan, so each pattern is backed by published sources, then tested and iterated with pre-2022 writing samples across genres against a LOT of AI writing. And maybe some LinkedIn content that triggered my spidey-senses.
What it isn’t
/human-eyes is a pattern detector, not an AI detector. It points out specific patterns in text that statistically give people the AI ick when used without intent.
It’s on GitHub. If you give it a try, tell me how it goes.