Deslopify

What it is

A skill file that rewrites AI‑sounding text so it reads like a person wrote it. It keeps the facts, refuses to add new ones, and when the source is vague it says so in a one‑line source‑gap note instead of inventing a story to fill the hole.

The rules live in two short files: a contract, and a catalogue of tells. Twenty‑four banned words, eleven banned phrases, a handful of banned constructions, and caps on the rest.

How it was made

Caps, not bans. Good writers use em dashes and lists of three, so the skill allows one of each per three hundred words. Stripping every device just produces a different kind of lifeless text.

A small Python checker reads the rules from the same file the model reads, so the two can never drift, and scores any text against them. A second script puts one sample through several models and scores each answer.

The sample was a LinkedIn post written to be as sloppy as possible. It walked in with thirty‑five tells.

The scorecard

Rule violations the checker found, lower is better. The first row is the input.

Sample		35
Claude Sonnet		0
Claude Haiku		1
GPT‑5.4 mini		3
Gemini 2.5 Flash		4

What I learned

Models obeyed the prohibition, do not add facts, four times out of four. They followed the conditional, say so when the source is vague, once. Prohibitions land; conditionals mostly do not.

Gemini passed several rules by find‑and‑replace. “In today’s fast‑paced world” became “In our fast‑paced world”. A word‑list checker cannot tell a rewrite from a swap, which is the most useful thing the checker taught me.

Inventing details tracked the size of the model, not the vendor.

The model that drafted the anti‑slop skill slipped an em dash and an aphorism into it. I caught both.

Credits

• Signs of AI writing, Wikipedia: the starting catalogue of tells.