deslop

name

deslop

description

De-slop pass for any text: detects and erases statistical fingerprints of AI writing (negative parallelism / "not X but Y", em-dash abuse, rule-of-three, false ranges, puffery vocabulary, uniform cadence, hedged both-sidesing) and rewrites into the target register — academic article, tweet, reddit post, email, blog, anything between. Use when the user says "deslop", "de-slop", "de-slop this", "remove the AI tells", "humanize this", "make this not sound like AI", or invokes /deslop. Also use before publishing any agent-drafted prose, including player-facing game copy (UI labels, item/skill descriptions, help text, hints). See deslop-player-facing-copy.mdc and deslop-markdown.mdc. Adapted from https://github.com/JuliusBrussee/skills.

Deslop

Strip every mark of AI writing from a text and make it good in its genre. Not "make it pass a detector" — make it read like a specific person with a specific point wrote it for a specific audience.

Why this is a loop, not a style guide

The worst tells — above all the "not X but Y" family — are not vocabulary mistakes. They are emergent properties of how LLMs generate text: preference tuning rewards balanced, contrastive, comprehensive-sounding framing, so the contrast move is baked into the model's priors. Two consequences drive this skill's architecture:

1. You cannot reliably see your own slop. The same priors that produce the pattern make it invisible on re-read. Detection must be mechanical — regex against a fixed catalog — never "does this look AI to me?"
2. Rewriting reintroduces slop. Ask a model to remove "it's not just X, it's Y" and it produces "this is less about X than Y" — the same move in a wig. So every rewrite gets re-scanned, and the loop runs until the scan is clean.

Workflow: Scan → Diagnose → Rewrite by meaning → Re-scan → (repeat) → Register check.

Phase 0: Fix the target

Before touching the text, establish:

• Genre and venue — academic article, tweet, reddit post, LinkedIn, email, blog, docs, marketing. If not stated and not obvious from the text, ask. Genre decides which tells are fatal and what "good" means; see references/voices.md.
• Audience and stance — who reads it, and what the author actually claims. Slop is what fills the space where a claim should be; you cannot remove it without knowing the claim.
• Constraints — length limits, required citations, house style.

Phase 1: Mechanical scan

Run the detection patterns from references/tells.md against the text. If the text is in a file (or you can write it to a temp file), run the grep commands in that reference literally — the catalog is written as runnable grep -Ein patterns. Otherwise apply each pattern by hand, line by line.

Produce a finding list: line/sentence, matched pattern, tell category. Also run the two structural checks that regex can't fully catch:

• Cadence: flag any run of 3+ consecutive sentences within ±4 words of the same length, and any paragraph where every sentence has the same shape (subject–verb–elaboration).
• Formatting: bold scattered through prose, emoji-decorated headers or bullets, "Term: definition" bullet lists, headers on a text too short to need them, a tidy intro–three-points–conclusion skeleton.

Report the findings to the user as a short table before rewriting (category, count, worst example). This is the diagnosis; the user should see what was wrong.

Phase 2: Rewrite by meaning, not by frame

Go finding by finding. The cardinal rule: never fix a pattern by paraphrasing the pattern. Fix it by deciding what the sentence actually asserts, then asserting that.

The "not X but Y" family — three-way triage

Every negative parallelism gets exactly one of these treatments:

1. The negation is a strawman (nobody believes X). Delete the X half entirely and assert Y directly, with whatever evidence the text has.
   • "It's not just a tool, it's a fundamental shift in how teams work" → "Teams that adopted it stopped holding standups within a month."
2. The contrast is real (people genuinely hold X). Then earn it: name who holds X, say concretely why Y beats it. A real contrast survives being made specific; slop doesn't.
3. The sentence asserts nothing (the contrast is decoration on an empty claim). Delete the whole sentence. Most cases are this one.

Banned escape hatches — these are the same move and count as new findings: "less about X than Y", "X matters, but Y matters more", "the real X is Y", "the question isn't X, it's Y", "X? Y." (rhetorical-question variant), and the em-dash variant "— not X, but Y".

Everything else

• Puffery and inflated vocabulary (pivotal, seismic, testament, tapestry, landscape, delve…): replace with the plain word, or with the concrete fact the puffery was hiding. "Plays a vital role in" → "does".
• Rule-of-three lists: keep the strongest item, cut the rest — unless all three carry distinct information, in which case keep them and break the rhythm (different lengths, different syntax).
• False ranges ("from X to Y"): if you can't name a meaningful midpoint between X and Y, it's not a range — name the two things or cut one.
• Hedged both-sidesing ("it's worth noting", auto-counterpoints, "while X, it's also true that Y"): commit. One opinion, stated, owned. A counterpoint stays only if the author genuinely concedes it.
• Uniform cadence: vary deliberately. Follow a long sentence with a short one. Fragments are legal. Don't apply a formula (alternating long/short is its own tell) — read the paragraph aloud and break wherever the rhythm is metronomic.
• Low specificity: replace "many companies" / "studies show" / "recent research" with the actual names, numbers, and dates — only from the source text, the conversation, or verifiable research you actually do. Never invent specifics. If the author needs to supply one, leave a marked placeholder: [ADD: which study?].
• Stock skeleton: kill throat-clearing openers ("In today's fast-paced world…"), summary conclusions ("In conclusion… Ultimately…"), and engagement-bait endings ("What do you think?"). Start where the point starts; stop when it's made.

What not to do — overcorrection is also slop

• No fake typos, forced slang, or manufactured "voice". Humanizer-tool output is its own genre of slop.
• Em dashes are not banned. Humans use them. The tell is density and the double-dash "— not X, but —" move. Budget: at most one em dash per ~150 words, never two in a sentence.
• Don't trade precision for personality in academic or technical text. There, de-slopping means cutting puffery and committing to claims — not adding attitude.
• Preserve the author's meaning, claims, and facts exactly. This is a style pass, not a content edit. Flag, don't silently fix, anything that looks factually wrong.

Phase 3: Verify loop

Re-run the full Phase 1 scan on your rewritten text. This step is not optional and not a formality — expect your own rewrite to contain new tells, because the model writing it has the same priors that created them. Fix and re-scan until a pass produces zero pattern hits and the cadence check passes. Cap at 4 passes; if a pattern survives 4 passes, rewrite that sentence from scratch starting from its bare claim ("what fact or opinion is this sentence for?").

Phase 4: Register check

Check the clean text against its genre profile in references/voices.md: right length, right formality, right person, genre-specific tells gone (e.g. on reddit: no bold, no bullet essay; in academic prose: no first-person hot takes added). Then the final test — read it aloud. Anywhere you wouldn't say it to the actual audience, rewrite that sentence.

Deliver: the rewritten text, plus a brief change log (categories fixed, counts, and number of verify passes it took).