deslop

Deslop: removing the signs of AI writing

LLMs have an identifiable writing style. Left unchecked, AI prose regresses toward the statistical mean: it smooths specific, unusual, verifiable facts into generic, positive, important-sounding filler. The result reads fluent but hollow - "slop." This skill is a field guide to catching and fixing those tells.

How to use this skill

There are two modes:

1. Gate mode (default, silent). When you are writing substantial prose, self-check the draft against the patterns below before presenting it. Don't announce that you're doing this; just produce clean output.
2. Audit mode (explicit). When the user gives you text and asks you to de-slop it, critique it, or check whether it sounds AI-written, scan against every category, then either (a) return an edited version, or (b) return a findings list with specific quoted offenders and fixes - match whatever the user asked for.

For a full audit of an external file, read references/full-checklist.md for the exhaustive pattern list with examples. The summary below covers the high-frequency offenders that catch ~90% of slop.

Critical mindset

• The patterns are signals, not crimes. Humans write some of these too (blogs, editorials, press releases). The presence of one phrase doesn't condemn a text; a cluster of them is the tell. Don't mechanically purge every "however."
• Fixing the surface tic is not the goal - fixing the underlying emptiness is. Deleting the word "underscores" while leaving a sentence that says nothing just makes the slop harder to detect. If a sentence only puffs up significance and carries no fact, cut the whole sentence, don't reword it.
• Specificity is the antidote. The core failure of slop is vagueness masquerading as importance. Replace "a revolutionary titan of industry" with "inventor of the first train-coupling device." When you can't add a real fact, delete the claim.

The high-frequency tells

1. Puffery: undue emphasis on significance and legacy