Detection patterns for the F*ck Slop scan. Patterns are written for `grep -Ein` (extended regex, case-insensitive, line numbers) so they can be run literally against a file:
```bash
grep -Ein -f /dev/stdin draft.txt <<'PATTERNS'
<paste patterns from a section below, one per line>
PATTERNS
```
When the text only exists in conversation, apply each pattern by hand. A match is a *finding*, not an automatic deletion — every finding goes through the Phase 2 triage in SKILL.md. Density matters: one em dash is nothing; one em dash plus a negative parallelism plus "delve" in the same paragraph is a verdict.
## 1. Negative parallelism — the "not X but Y" family
The highest-priority category. LLMs reach for the negation-then-assertion move roughly once a paragraph; humans use it occasionally and deliberately. It is an emergent generative habit, so expect it to reappear in paraphrased form after every rewrite pass — a targeted rescan may help after a substantial rewrite.
we'?re not (just )?(talking about|looking at|dealing with)
gone are the days
(here|this)'?s the (thing|kicker|catch|twist)
```
Rhetorical-question variant (regex-resistant; check by hand): a one-line question immediately answered by a one-word or one-clause sentence. *"The result? Chaos."* / *"Sound familiar?"*
## 2. Puffery and inflated vocabulary
Single words that spike in LLM output. Each is fine in isolation; two or more per page is a finding. The fix is the plain word or the concrete fact the word was hiding.
let'?s (dive|unpack|explore|take a (look|closer look))
whether you('re| are) [^.;]{2,60} or
```
## 4. False ranges and rule-of-three
**False range** — a "from X to Y" with no actual spectrum between X and Y:
```
from [^.;]{3,50} to [^.;]{3,50}
```
Triage by hand: if you can name a meaningful midpoint, it's a real range and stays. If X and Y are just two loosely related examples, name them plainly or cut one.
**Rule of three** — LLMs default to triplets to make thin analysis look thorough. Regex only catches the simplest shape; check lists by hand too.
- **Headers on short texts** — section headers on anything under ~400 words.
- **The tidy skeleton** — intro that previews three points, three matched sections, conclusion that restates them. Resolves too neatly; real writing has loose ends.
- **Numbered lists where a paragraph would do.**
- Curly quotes/apostrophes in a context where the author types straight ones (mixed within one text is the stronger tell).
- **Uniform sentence length** (the single strongest current tell): a run of 3+ consecutive sentences within ±4 words of each other, paragraph after paragraph of 18–24-word sentences. Optional diagnostic on a file (not a completion gate):
Human prose mixes 4-word sentences with 30-word sentences. Variance should be obvious at a glance.
- **Uniform sentence shape**: every sentence opens subject-first; no fragments, no questions, no inversions.
- **Uniform paragraph length**: every paragraph 3–4 sentences.
- **Low specificity**: "many companies", "studies show", "experts agree", "recent research", "various factors" — generic where a human who knew the material would name names, numbers, dates. (Fix only with real specifics; never invented ones.)
- **No friction**: nothing colloquial, no aside, no opinion held without a softener, nothing that risks being disagreed with.
## 7. Genre-specific instant tells
Covered in detail in [voices.md](voices.md); the headline items:
