lyndonkl/skills/slop-detector

Slop Detector

Table of Contents

• The 10 signatures
• Workflow
• Worked example
• Guardrails

Related skills: Called by the Editor voice pass. Consumes hedge-cluster count from hedge-detector (S8). Emits the "Slop signatures" subsection.

The 10 signatures

Fixed list. Each either clean or flagged with the offending span.

| # | Signature | Detection | |---|---|---| | S1 | Meta-framing opener | First paragraph contains In this post, This article, We will explore, Let's dive into, Today we'll look at | | S2 | List-carrying-argument | Any bulleted list where the argument collapses if bullets are removed. Test: does the prose still stand without the list? | | S3 | Zombie nouns (Sword) | >3 nominalizations per 100 words (suffixes: -ation, -ity, -ment, -ence on abstract nouns) | | S4 | Generic examples | "a company" / "a model" / "a user" with no specific name, scale, dataset | | S5 | No first-person | Zero I, my, we-as-me in a >800-word reflective essay | | S6 | Prompt residue | Let's break this down, To summarize, In conclusion, Key takeaways, Let me explain | | S7 | Outline-shaped paragraphs | >60% of paragraphs follow same syntactic shape: topic → 3 supporting sentences → transition | | S8 | Hedge cluster | ≥2 epistemic-weakness hedges within 50 words (from hedge-detector) | | S9 | Buzzword stuffing | ≥3 terms from {game-changer, paradigm shift, under the hood, delve, unpack, dive into} in a single draft | | S10 | Flattened uncertainty | Any small-N caveat that appears in corpus/drafts/notes/ but was removed in the submitted draft (requires notes; else skip this signature) |

Workflow

Slop scan draft D:
- [ ] Step 1: For each signature, run detection rule
- [ ] Step 2: Mark each signature as clean | flagged (with quote)
- [ ] Step 3: Tier-1 signatures: S1, S2, S6 (generic framing + prompt residue)
- [ ] Step 4: Tier-2 signatures: S3, S4, S5, S7, S9
- [ ] Step 5: Emit the slop signatures subsection with each labeled clean/flagged

S3 nominalization scoring

Count suffix hits (-ation, -ity, -ment, -ence, -ness, -ance) on abstract nouns per 100 words. >3 = flag. Example: "provides analysis of" → nominalized; "analyzes" → active.

S4 generic-example rule

Flag an example if it uses only generic pronouns / nouns without a specific anchor:

• "A company might use this" → flag.
• "At Google in 2024, Chen et al. used this" → clean.

S7 outline-shape rule

Parse paragraphs; count those with the shape:

• Sentence 1: topic statement
• Sentences 2–4: three supporting sentences
• Last sentence: transition

60% of paragraphs following this shape → the draft reads like an AI-generated outline expanded.

Worked example

Draft fragment:

In this post, we'll explore why RAG beats fine-tuning.

First, let's define RAG. It's a technique where models retrieve documents before generating. A company might use RAG for their customer service chatbot.

Second, fine-tuning involves training. A team might fine-tune to adapt style.

Third, RAG has benefits. Fine-tuning has drawbacks. It could be argued that hybrid works.

To summarize, both approaches have merit.

Detections:

• S1: flagged ("In this post, we'll explore").
• S2: flagged (argument carried by "First / Second / Third" list-in-prose).
• S3: zombie-noun check — "technique", "documents", "benefits", "drawbacks" — borderline. Not flagged yet.
• S4: flagged ("A company might use RAG", "a team might fine-tune") — no specifics.
• S5: clean (has "we").
• S6: flagged ("To summarize").
• S7: flagged (each paragraph: topic + supporting + transition).
• S8: weakness hedges — "It could be argued" — cluster check: just 1, not a cluster (yet).
• S9: buzzword — "explore" is close. Not flagged yet (1 term).
• S10: skipped (no notes dir).

Output: 5 signatures flagged (S1, S2, S4, S6, S7). Tier-1: S1, S2, S6 = 3 tier-1 slop violations.

Guardrails

1. Each signature has a concrete detection rule. No "feels slop."
2. Quote the offending span. Don't just say "S1 triggered" — quote the opener.
3. Signatures are additive, not exclusive. A draft can trip 8 signatures and still be revise-able; Editor's must-not #13 sets the no-go threshold.
4. S10 requires a notes dir; skip quietly if absent.
5. Don't double-count with hedge-detector. Hedge clusters flow from hedge-detector into S8 as an input, not a separate scan here.
6. S5 (no first-person) — reflective essays only. How-to / methodology posts may legitimately lack "I."

Quick reference

• 10 fixed signatures, deterministic detection.
• Tier-1: S1, S2, S6. Tier-2: the rest.
• Consumes hedge-detector cluster count as S8 input.