86 lines
4.6 KiB
Markdown
86 lines
4.6 KiB
Markdown
|
---
|
|||
|
|
name: fuck-slop
|
|||
|
|
description: >
|
|||
|
|
De-slop pass for any text: detects and erases statistical AI writing fingerprints
|
|||
|
|
(negative parallelisms, em-dash abuse, rule-of-three, false ranges, puffery, uniform cadence)
|
|||
|
|
and rewrites prose to match target genre registers. Use when asked to "fuck slop", "deslop",
|
|||
|
|
"remove AI tells", "humanize", or before publishing agent-drafted prose.
|
|||
|
|
concerns: [style]
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
# F*ck Slop
|
|||
|
|
|
|||
|
|
Strip every mark of AI writing from a text and make it authentic to its genre. Not "make it pass an AI detector" — make it read as if a specific human with a specific point wrote it for a specific audience.
|
|||
|
|
|
|||
|
|
## The Core Problem
|
|||
|
|
|
|||
|
|
The worst tells — especially **"not X but Y"** negative parallelisms — are emergent properties of preference tuning rewarding contrastive, comprehensive-sounding framing.
|
|||
|
|
1. **You cannot reliably see your own slop**: Detection must be mechanical (regex against a fixed catalog), never subjective impressions.
|
|||
|
|
2. **Rewriting reintroduces slop**: Rewriting "it's not just X, it's Y" often outputs "this is less about X than Y" — the same tell in a wig. The scan must loop until clean.
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## Phase 0: Setup and Protocols
|
|||
|
|
|
|||
|
|
1. Review governing catalogs:
|
|||
|
|
- [references/tells.md](references/tells.md) for runnable `grep -Ein` patterns across all diagnostic categories.
|
|||
|
|
- [references/voices.md](references/voices.md) for target register profiles (academic, tweet, reddit, email, blog, docs, marketing).
|
|||
|
|
2. Rule: Never fix a pattern by paraphrasing the pattern. Fix by asking what the sentence asserts, then asserting that directly.
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## Phase 1: Ingest and Check
|
|||
|
|
|
|||
|
|
1. **Establish the Target**:
|
|||
|
|
- **Genre & Venue**: Academic article, technical documentation, blog, tweet, email, Reddit post.
|
|||
|
|
- **Audience & Stance**: Who reads it, and what does the author specifically claim? Slop fills the space where a claim should be.
|
|||
|
|
- **Constraints**: Word limits, house style, required terminology.
|
|||
|
|
2. **Mechanical Scan**:
|
|||
|
|
- Run the regex catalog from [references/tells.md](references/tells.md) against the text.
|
|||
|
|
- Flag structural rhythm failures:
|
|||
|
|
- **Cadence**: 3+ consecutive sentences within $\pm 4$ words of equal length.
|
|||
|
|
- **Formatting**: Bold scattering through prose, emoji bullet spam, symmetrical triads.
|
|||
|
|
- Summarize findings into a concise diagnostic list before rewriting.
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## Phase 2: Processing (Core Logic)
|
|||
|
|
|
|||
|
|
### Step 1: Negative Parallelism Triage
|
|||
|
|
Every "not X but Y" pattern receives one of three treatments:
|
|||
|
|
1. **Negation is a strawman** (nobody believes X): Delete X entirely; state Y directly with concrete evidence.
|
|||
|
|
2. **Contrast is genuine** (people hold X): Name who holds X and show concretely why Y wins.
|
|||
|
|
3. **Sentence asserts nothing** (decoration over empty claim): Delete the entire sentence.
|
|||
|
|
|
|||
|
|
*Banned escape hatches*: "less about X than Y", "X matters, but Y matters more", "the question isn't X, it's Y".
|
|||
|
|
|
|||
|
|
### Step 2: Vocabulary & Structure De-Slopping
|
|||
|
|
- **Puffery Tokens**: Replace inflated words (`pivotal`, `seismic`, `testament`, `tapestry`, `delve`, `multifaceted`) with plain verbs or concrete metrics.
|
|||
|
|
- **Rule-of-Three Lists**: Keep the single strongest item and discard filler. If keeping all three, break the symmetrical syntax.
|
|||
|
|
- **False Ranges**: ("from X to Y"): If no meaningful midpoint exists, name the two separate entities or cut one.
|
|||
|
|
- **Hedged Both-Sidesing**: ("it is worth noting", auto-counterpoints): Commit to a clear stance. Concede counterpoints only if genuine.
|
|||
|
|
- **Cadence Variance**: Mix short punchy sentences (3–6 words) with compound analytical statements. Break metronomic rhythm.
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## Phase 3: The System One Balance Pass
|
|||
|
|
|
|||
|
|
- **Preserve Ground Truth**:
|
|||
|
|
- Do NOT alter author claims, empirical numbers, code identifiers, or citations. This is a stylistic de-slop pass, not content censorship.
|
|||
|
|
- **Avoid Overcorrection**:
|
|||
|
|
- Do NOT inject fake typos, manufactured slang, or artificial cynicism ("humanizer" slop).
|
|||
|
|
- Em dashes are permitted when used naturally (maximum one per 150 words; never two in a sentence).
|
|||
|
|
- In technical and academic registers, prioritize clarity and precision over informal attitude.
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## Phase 4: Output Execution & Verification
|
|||
|
|
|
|||
|
|
1. **Verification Loop**:
|
|||
|
|
- Re-run the Phase 1 scan on the rewritten text.
|
|||
|
|
- Fix and re-scan until zero regex hits occur (maximum 4 passes; if a sentence fails 4 passes, discard and rewrite its bare assertion from scratch).
|
|||
|
|
2. **Register Verification**:
|
|||
|
|
- Verify formatting and tone match [references/voices.md](references/voices.md).
|
|||
|
|
3. **Delivery**:
|
|||
|
|
- Output the clean, de-slopped text.
|
|||
|
|
- Provide a concise changelog table: Tell category, occurrences caught, and passes required.
|