avoiding-ai-slop

Detects and eliminates AI-slop from text, prose, and image prompts using a 76-rule structural pattern corpus, a 237-term tiered ban lexicon, and an image-slop negative-prompt pack, with a slopcheck.py

SKILL.md

GitHub

Avoiding AI Slop

Structural pattern rules beat word blacklists: vocabulary decays once publicized ("delve" usage dropped after early-2024 mockery — Kobak, arXiv 2406.07016), but sentence shapes persist across model generations. This skill enforces 76 pattern rules + a tiered lexicon in three modes.

Quick-Start Decision Tree

What's the task?
  ├─ WRITING new prose ──────────── Mode A: apply the constraint digest below, silently
  ├─ REVIEWING/de-slopping a draft ─ Mode B: lint it, then rewrite every hit
  ├─ CI / pre-publish gate ───────── Mode B: scripts/slopcheck.py --scan, exit-code wired
  ├─ IMAGE prompt or image review ── Mode C: references/image-slop.md pack + checklist
  └─ Thresholds feel wrong ───────── Calibrate: --calibrate on 10 known-good docs

Usage Examples

Example 1 — "De-slop this LinkedIn draft." → Run python3 scripts/slopcheck.py --scan draft.md. Rewrite every BAN hit using the fix: guidance per rule; adjudicate FLAG hits in context; deliver the rewrite, not a description of it.

Example 2 — "Write a blog post about our n8n migration." → No slop-review was requested; Mode A applies silently. Draft under the constraint digest. Before delivering, self-scan mentally against the eight pattern families — no negative parallelism, no tailing significance clauses, no editorial framing.

Example 3 — "Does this read like AI wrote it?" → Scan, report the specific rule IDs that fired with line numbers, and say what the hits are — evidence, not vibes. Never extend this to an authorship claim about a third party; the corpus detects style, not provenance.


Mode A — Generation-time constraints (the digest)

Apply when writing any outward-facing prose. This digest is the injection payload; do NOT paste the full lexicon into a generation context — long ban-lists trigger ironic-rebound usage and burn context. Patterns generalize.

Contrast templates — never:

  • "It's not X, it's Y" / "not just X, but Y" reframes (CT-01/02)
  • "The result? ..." fake Q&A cadence; "Here's the thing" pivots (CT-04/05)
  • Rule-of-three triplet lists as default rhythm; em-dash saturation (CT-06/07)

Tailing clauses — never:

  • ", highlighting/underscoring/showcasing/reflecting..." significance tails (TL-01)
  • "stands as a testament to", "plays a crucial role in", "cemented its status", "lasting impact", "marks a significant shift" (TL-03..09)
  • End the sentence at the fact. Prove significance in its own sentence or cut.

Editorializing — never:

  • "It's important/worth noting", "In today's fast-paced world", "Whether you're a X or a Y", "Let's dive in", "In conclusion" scaffolds (ED-01..07)

Hedging — never:

  • "Experts argue" / "widely regarded as" without a named, cited source (HG-01/02)
  • "a plethora/myriad of", "has the potential to", stacked modals (HG-04/06/07)

Promo puffery — never:

  • "nestled in the heart of", "rich tapestry", "game-changer", "unlock/ unleash/harness the power of", "navigate the complexities", "evolving landscape", "embark on a journey", "foster a culture of" (PR-01..12)

Structure:

  • No emoji as bullets/headers. No bold-leadin bullet scaffolds as default. No "Introduction/Conclusion/FAQ" headers on short pieces. Vary paragraph and sentence length — kill the 2–3-sentence metronome (ST-01..10).

Artifacts — never ship:

  • "As an AI model", knowledge-cutoff disclaimers, "I hope this helps", "[insert X]" placeholders, oaicite/[cite: N]/turn0search debris (AR-01..07)

Fiction:

  • No chills down spines, breaths they didn't know they were holding, eyes widening in surprise, voices barely above whispers, air thick with anything, "couldn't help but", "little did they know" (FX-01..12)

Top-40 word bans (Tier 1, all inflections): delve, tapestry, multifaceted, synergy, holistic, embark, unwavering, unparalleled, meticulous, intricacies, pivotal, garner, glean, bolster, spearhead, showcase, supercharge, unleash, revolutionize, transformative, groundbreaking, breathtaking, captivating, mesmerizing, bustling, indelible, ever-evolving, fast-paced, beacon, cornerstone, culmination, epitome, quintessential, veritable, undoubtedly, invaluable, paramount, utilize, whilst, game-changer. Full 117-term list + replacements: references/lexicon.yaml.

Register calibration: tiers assume technical/MSP-consulting writing. robust, significant, end-to-end, turnkey, roadmap, framework are Tier 2/3 (context/density), never banned — they're working vocabulary in engineering, stats, and MSP delivery.

Mode B — Lint-time

python3 scripts/slopcheck.py --self-test          # corpus integrity (run after any edit)
python3 scripts/slopcheck.py --scan draft.md      # BAN=error, FLAG=warn, DENSITY vs thresholds
python3 scripts/slopcheck.py --scan draft.md --overrides overrides.yaml --allow allow.yaml

Exit codes: 0 clean · 1 findings · 2 config error. DENSITY tiers are skipped under 150 words (per-1000 rates are noise on small samples). Matching is inflection-aware (delve→delving, tapestry→tapestries, game-changer→game changers); -ness/-ation variants are out of scope by design.

Calibration workflow (do once before trusting DENSITY tiers):

python3 scripts/slopcheck.py --calibrate good1.md good2.md ... > overrides.yaml

Feed ~10 documents you consider clean; any threshold they breach gets a suggested replacement (observed max × 1.25). Shipped thresholds are reasoned defaults, not measurements — calibrate before wiring into CI.

Allowlist: per-project exemptions in allow.yaml:

allow: [ST-05, "framework", "key takeaways"]   # rule ids and/or terms

Mode C — Image prompts and review

references/image-slop.md: six negative-prompt groups (skin, lighting, grade, composition, tech-cliché motifs, prompt-boilerplate bans) — pick groups per job, don't stuff all of them. Then the 12-point review checklist; any "no" = regenerate. Icon layer belongs to purging-icon-slop; brand tokens to the Abstrekkt/design-brief gates. No duplication here.


Reference Map (progressive disclosure)

Need Open
A specific rule's regex, examples, fix references/patterns.yaml (76 rules, 8 categories, BAN/FLAG/DENSITY)
Full ban list, context notes, ~95 replacements references/lexicon.yaml (117 ban / 70 flag / 50 density)
Image negatives + audit checklist references/image-slop.md
Lint, calibrate, allowlist mechanics scripts/slopcheck.py --help
Trigger evals evals/evals.md
What changed 2.0→5.0 CHANGELOG.md

Refresh Protocol (quarterly)

  1. --self-test must pass before and after any corpus edit.
  2. Re-scan 3 recent frontier-model outputs; add new tells as rules (patterns preferred) or terms (exactly one tier each).
  3. Retire vocabulary that has stopped appearing — decay is expected; note retirements in CHANGELOG.
  4. Re-run --calibrate if your own writing register shifted.
  5. Bump versions independently: SKILL.md, patterns.yaml meta.version, lexicon.yaml meta.version, slopcheck __version__ — components version separately, no forced sync.

Self-Verification (before delivering de-slopped or new prose)

  1. Scan ran (or digest applied) — zero unadjudicated BAN hits
  2. FLAG hits reviewed in context, not blanket-rewritten
  3. Paragraph rhythm varies (ST-10 — the one check regex can't do)
  4. No artifact-tier leakage (openers, closers, placeholders, citation debris)
  5. Rewrites carry the content forward — de-slopping never deletes facts
  6. Images: checklist run, any "no" regenerated

CLAIMS

  • Pattern rules outlive word lists; the delve-decay evidence is the load-bearing argument (Kobak, arXiv 2406.07016)
  • Density thresholds shipped here are [GENERATED] defaults — uncalibrated until run against the user's own clean corpus
  • Style detection is not authorship detection; this skill never supports "prove X used AI" claims

COUNTEREXAMPLE

  • Aggressive pattern-banning can produce stilted avoidance-voice — a different, detectable slop; Mode A's restraint (digest, not full lexicon) is the mitigation
  • A human who genuinely writes "testament to" gets flagged; FLAG-tier adjudication exists precisely because the corpus has false positives

CONTRADICTIONS

  • "Ban these words" vs "don't paste ban-lists into prompts" (ironic rebound) — resolved by splitting generation-time digest from lint-time full corpus, but the tension is real
  • Register calibration says significant is fine in stats while Tier-3 caps it — a stats-heavy doc will warn on legitimate usage until calibrated