Prose Deslop

Prose Deslop

Eliminate predictable AI writing patterns from prose. Loop: Write → Review → Revise until the rubric passes.

Modes

Identify which job you are doing before touching the text:

• Generate: writing new prose. Apply the Core Rules from the first sentence. Don't write slop and then clean it. Specifics must come from the user's brief or source material; where they're missing, use a clearly marked placeholder ("[X% — author to confirm]") or ask.
• Edit: rewriting someone else's draft. Preserve the author's meaning, facts, and voice; remove only the patterns. Never delete information to satisfy a rule, and never add information to satisfy one: no invented numbers, names, anecdotes, or examples. When a rule would change what the author is saying, leave the sentence and flag it instead.
• Review: auditing without rewriting. Produce a report (format below), don't change the text unless asked.

Workflow

1. Write/Edit following the Core Rules below. Load the reference files when working on a draft of any length.
2. Review the draft against the Quick Checks, sentence by sentence. Triage findings by the severity tiers in references/structures.md: fix critical-tier patterns (throat-clearing, binary contrasts, structural issues) before cosmetic ones.
3. Revise and score against the rubric. Below 35/50: revise again.
4. Anti-overcorrection check: the result must stay informative and varied, not terse. If the revision lost facts, numbers, or voice, restore them. See Example 11 in references/examples.md.

Review Report Format

When asked to review text (not rewrite it), report findings like this:

Slop Review: <name>
Verdict: low | moderate | high | severe

Critical:
- "Here's the thing" (line 3) - throat-clearing, cut
- "isn't about X. It's about Y" (line 3) - binary contrast, state Y directly
High:
- "The implications are significant" (line 7) - vague declarative, name the implication
Medium/Low:
- "leverage", "every... always" (line 9) - jargon, lazy extremes

Rubric: Directness 4/10, Rhythm 6/10, Trust 5/10, Authenticity 4/10, Density 5/10 = 24/50 -> revise

Quote the exact offending words with their location, name the pattern, give the fix direction. Severity tiers and thresholds are in references/structures.md.

Core Rules

1. Cut filler phrases. Remove throat-clearing openers, emphasis crutches, business jargon, buzzwords, hedge clusters, wordy constructions, and crutch adverbs. See references/phrases.md.
2. Break formulaic structures. No binary contrasts ("not X, it's Y"), negative listings, dramatic fragmentation, rhetorical setups, false agency. See references/structures.md.
3. Use active voice. Every sentence needs a human subject doing something. No inanimate objects performing human actions ("the complaint becomes a fix").
4. Be specific — with real specifics only. No vague declaratives ("The reasons are structural"). Name the specific thing if it exists in the source material or the author supplied it. Never invent numbers, names, quotes, companies, or anecdotes to satisfy this rule; if the specific is missing, flag the vague sentence and ask the author for the detail (see Example 13 in references/examples.md). A vague true sentence beats a vivid fabricated one. No lazy extremes ("every," "always," "never") doing vague work.
5. Put the reader in the room. No narrator-from-a-distance voice. "You" beats "People." Specifics beat abstractions.
6. Vary rhythm. Mix sentence lengths. Two items beat three. End paragraphs differently. Avoid em dashes by default (an AI tell when frequent); an occasional deliberate one is acceptable.
7. Trust readers. State facts directly. Skip softening, justification, hand-holding, and meta-commentary about what the text will do.
8. Cut quotables. If it sounds like a pull-quote, rewrite it.
9. Lead with content. No opening paragraphs that are pure preamble, no closings that restate without adding.

These rules are defaults, not laws. When a rule conflicts with the genre, the author's voice, or the facts, the Context Matters section below wins.

Quick Checks

Before delivering prose:

• Crutch adverbs ("really," "very," "actually," "genuinely")? Cut them; keep only adverbs that change meaning.
• Any passive voice? Find the actor, make them the subject.
• Inanimate thing doing a human verb ("the decision emerges")? Name the person.
• Statement starts with a Wh- word ("What makes this hard is...")? Restructure. Genuine questions are fine.
• Any "here's what/this/that" throat-clearing? Cut to the point.
• Any "not X, it's Y" contrasts? State Y directly.
• Any "delve", "leverage", "utilize", "navigate the complexities"? Replace with plain words.
• Hedge cluster ("could potentially", "it seems that")? One precise statement.
• Wordy construction ("in order to", "due to the fact that")? Simplify.
• Three consecutive sentences match length? Break one.
• Paragraph ends with punchy one-liner? Vary it.
• Em dashes clustered (several in a short span, or more than ~1 per 500 words)? Replace with commas or periods.
• Every paragraph starts with a transition word? Restructure.
• Vague declarative ("The implications are significant")? Name the specific implication — or flag it if the specific isn't in the source.
• Any number, name, quote, or anecdote not present in the source or supplied by the author? Remove it or mark it as a placeholder. Never ship invented specifics.
• Narrator-from-a-distance ("Nobody designed this")? Put the reader in the scene.
• Meta-joiners ("The rest of this essay...", "In this article we will...")? Delete.

Scoring

Rate 1-10 on each dimension:

Calibration anchors (apply per dimension):

• 2 — the flaw appears in most paragraphs (e.g., nearly every paragraph opens with throat-clearing, or sentence lengths are metronomic throughout).
• 5 — the flaw appears a few times in the piece; a casual reader might not notice, an editor would.
• 8 — at most one borderline instance; nothing an editor would mark.
• 10 — reserve for text you could not improve on that dimension.

Below 35/50: revise. The score is a forcing function for a second pass, not a measurement — always re-read the flagged lines rather than trusting the number.

Context Matters

This section takes precedence over the Core Rules and Quick Checks. The rules target overuse and unconscious repetition — the AI default — not the mere existence of a pattern. Apply judgment:

• Academic writing legitimately hedges more than blog posts.
• Legal, regulatory, and compliance text needs precision over punch; hedges ("may", "to the extent that") and passive constructions are often deliberate and required. Do not apply this skill to such text unless the user explicitly asks; even then, flag rather than rewrite anything that could change legal meaning.
• A skilled writer's occasional em dash or adverb is fine; the rules exist because AI defaults to them constantly.
• When a rule fight breaks the author's voice, meaning, or factual accuracy, voice, meaning, and accuracy win.

Examples

See references/examples.md for before/after transformations.

Output Expectations

This skill moves text from "obviously AI" to "neutral, direct prose." It cannot inject genuine voice, lived detail, or domain knowledge; those come from the author — never fabricate them to satisfy the "be specific" rule. Treat output as a draft requiring human review before publication.

Passing the rubric is the floor, not the finish. Deslopped prose can be clean and still forgettable. Voice comes from things the rubric does not score: a real point of view the writer is willing to commit to, concrete detail only this author would know, and rhythm chosen for effect rather than merely varied to avoid the metronome. When the slop is gone and the writing still reads flat, the fix is not another deslop pass. It is the author adding something only they can supply. Say so rather than scrubbing the text further.

A note on AI detectors

People often arrive at "humanize my text" wanting to pass an AI detector. This skill is not a detector-evasion tool, and it should not be sold as one. Removing AI tells does tend to lower detector scores, because the tells and the statistical signals overlap, but detectors are unreliable in both directions (they false-positive on human writing and flag non-native English at higher rates), so a passing score is not the goal and was never trustworthy. The aim here is prose that reads well to a person. If the request is to disguise AI-written work as human for a grade or other authorship claim, that is an integrity problem the skill does not solve; help the author write, not hide.