| name | pprose-de-slop |
| description | Remove AI-writing tells and formulaic LLM prose while preserving meaning and voice; modifies the target. Use when asked to de-slop, de-slopify, humanize, remove AI tells, or make writing sound less AI-generated or less machine-like; also use, unprompted, when text being written, edited, or reviewed reads machine-generated (an AI-heavy draft, pasted chatbot output). Fixes hollow emphasis, canned transitions and conclusions, boilerplate openers, inflated significance, false agency, negative parallelism, dramatic fragments, marketing register, chat-artifact leakage, and reflexive three-item lists. |
Remove AI-Writing Tells
This is a focused apply skill: it edits prose to remove documented LLM-register tells
without installing a different authorial voice.
It may modify the target document.
Before acting, read the bundled
ai-prose-corrections.md in full.
The reference is part of this skill; the pprose CLI is not required.
In the edit ladder this skill sits between pprose-common-edit (the universal
structure-and-formatting tier) and pprose-copy-edit; both pprose-copy-edit and
pprose-full-edit include this audit.
Apply de-slop wherever prose is drafted or edited, not only on request; load the
reference’s Drafting Directives while writing new text.
Reserve the broader tiers for when a fuller edit is wanted.
This skill improves prose quality.
It does not promise detector evasion, hide authorship, or manufacture signs of human
composition.
Inputs
- One or more prose documents or clearly identified passages.
- Optional audience, genre, voice sample, and tolerance for heavier rewriting.
Steps
- Read
references/ai-prose-corrections.md completely before editing.
- Read the full target and identify its purpose, audience, genre, intended voice, and
local authoring rules.
- Audit in two passes:
- Find lexical and structural patterns the reference marks for correction.
- Inspect attention flags for repetition or density; do not treat them as banned
words or constructions.
- Rewrite confirmed tells by naming the actor, claim, relationship, mechanism,
quantity, or consequence the original obscured.
- Preserve factual meaning, claim strength, citations, technical terms, and useful
structure. Keep intentional rhetoric when it passes the reference’s licensing test or
a genre convention requires it.
- Do not mechanically swap synonyms or add typos, slang, contractions, anecdotes,
quirks, or unevenness to simulate a person.
Do not make claims about detector scores or detector evasion.
- Follow the target project’s formatting and validation rules.
If the target is Markdown, run its configured formatter when one exists.
- Re-read the result and diff for changed meaning, flattened voice, new repetition,
unsupported specificity, or awkward rhythm.
Output
Report the changed files or passages, the main tell classes corrected, and any flagged
construction retained because it carried information or fit the genre.
