antislop-expert

name: antislop-expert description: "This skill should be used when the user asks to 'check for slop', 'audit this text', or 'is this authentic?'." model: sonnet

AntiSlop Expert (The Authenticity Guard)

This is not just a spellchecker. It is a forensic linguistics engine designed to root out "LLM-ese" and enforce {{voice_dna}}.

┌─────────────────────────────────────────────────────────────────┐
│  STANDALONE (always works)                                      │
│  ✓ Slop Detection: Flags "Delve", "Tapestry", "Game-changer".   │
│  ✓ Voice Match: Enforces your specific Tone and Sentence length.│
│  ✓ Bilingual Audit: Special rules for English & French nuance.  │
├─────────────────────────────────────────────────────────────────┤
│  SUPERCHARGED (connect ~~browser / local_files)                 │
│  + Bulk Audit: Scan entire folders of content at once.          │
│  + Competitor Benchmark: Compare your "Humanity Score" vs them. │
│  + Live Coaching: Real-time feedback as you type (in editor).   │
└─────────────────────────────────────────────────────────────────┘

🛠 Triggers

• "Audit this text for AI slop"
• "Make this sound more like me"
• "Check if this is written by ChatGPT"
• "Humanize this draft"

🛠 Agent Instructions

Before Auditing

1. Load Voice DNA: Read ${CLAUDE_PLUGIN_ROOT}/data/voice-dna.json. You CANNOT judge authenticity without this baseline.
   • Look for: sentence_length_variance, prohibited_words, tone_keywords.
2. Load Audience Profile: Read ${CLAUDE_PLUGIN_ROOT}/data/icp.json.
   • Purpose: Verify if the language complexity matches {{icp.reading_level}}.

1. Language Detection & Rules

🇬🇧 English Mode (Detects "Corporate Slop")

Markers to Hunt: (Refs: lexicon_slop_en.md, structural_patterns_en.md, stylometry_en.md)

• Hollow Adjectives: "Robust", "Seamless", "Cutting-edge", "Game-changing", "Revolutionary".
• Passive Voice abuse: "Mistakes were made", "It has been decided".
• Latinate Stacking: "Utilization of leverage for optimization".
• Structure: The "Introduction-Body-Conclusion" sandwich often used by LLMs.
• Vague Quantifiers: "Many", "Significant", "Various", "A number of".
• Voice Conflict: Any word in {{voice_dna.forbidden_words}}.

English Remediation Rules: (See full guide: ./references/remediation_rules_en.md)

• Anglo-Saxon > Latinate: Use "buy" not "purchase", "use" not "utilize".
• Active Verbs: "We failed" not "Failure occurred".
• Kill the Adverbs: "He ran" not "He ran quickly".
• Specifics: Replace "significant savings" with "$10k saved".

🇫🇷 French Mode (Detects "Langue de Bois")

Markers to Hunt: (Refs: lexicon_slop_fr.md, motifs_structurels_fr.md, stylometrie_fr.md)

• Corporate Jargon: "Synergie", "Holistique", "Travailler en mode projet", "ADN", "Focus".
• Nominalization: "La mise en place de l'optimisation" (vs "Optimiser").
• Empty Phrases: "Dans un monde en mutation", "Il est important de noter".
• Typography: Missing non-breaking spaces before (:;?!), wrong quote marks (" " vs « »).
• Passive & On: "Il a été décidé", "On a fait" (imprecise).

French Remediation Rules: (Voir guide complet : ./references/regles_remediation_fr.md)

• Verbes d'action: "Nous avons construit" vs "La construction a été faite".
• Chasse au "De": Éviter les chaînes de compléments ("La gestion de la mise en œuvre de...").
• Concret: Remplacer "solution globale" par ce que c'est vraiment (logiciel, tournevis, méthode).
• Typographie: Force French typography rules (espaces insécables, guillemets français).

1b. Slop Score Calculation

Every audit produces a Slop Score (0–100) defined as:

Slop Score = (Slop Tokens / Total Tokens) × 100
           + (Structural Patterns detected × 5)
           + (Voice conflicts / Total sentences × 20)

Calibration:

• 0–10: Clean. Minimal intervention needed.
• 11–25: Light slop. Targeted replacements sufficient.
• 26–50: Moderate. Structural rewrite recommended for at least 30% of the text.
• 51–75: Heavy. Version B full restructure required.
• 76–100: Unrecoverable. Rewrite from scratch using Voice DNA only.

Show the score with its components in the audit output:

"Slop Score: 34/100 — Slop tokens: 12%, Structural patterns: 2 (IBC sandwich + passive abuse), Voice conflicts: 18%"

2. Interaction Workflow

Slop Score Formula (0–100)

The score is not a gut feeling — it is calculated:

Score interpretation:

• 0–25: Human voice — minor adjustments needed
• 26–50: AI-influenced — remediation recommended
• 51–75: AI-generated — structural rewrite needed
• 76–100: Unambiguous LLM output — discard and rewrite from scratch

Mode 1: The Audit (Forensic Scan)

1. Scan against {{voice_dna}} and the Language Specific Markers above.
2. Calculate Slop Score using formula above — show the breakdown, not just the number.
3. Highlight specific offenses with category label.

Mode 2: The Remediation (Rewrite)

1. Strip adjectives/adverbs.
2. Inject {{voice_dna.idioms}}.
3. Output two versions:
   • Version A (Polish): Word-level fixes — replaces slop words, fixes typography, cuts adverbs. Same structure.
   • Version B (Humanize): STRUCTURAL rewrite. Must change:
     • The point of view (who is speaking and to whom)
     • The proof style (add a specific number, example, or story)
     • The sentence length pattern (match {{voice_dna.sentence_variance}})
     • The opening (never start with "In today's..." or "In a world...") Version B is not Version A with different words. It is a fundamentally different piece that communicates the same idea.

Mode 3: The Sparring (Challenge)

Ask 3-5 probing questions to challenge vague claims.

• EN: "You say 'significant efficiency gains' — exactly how many hours per week?"
• FR: "Vous parlez de 'synergie' — quel est le gain financier concret ?"

📝 Output Format

Produce the report in the same language as the input text.

# 🛡️ AntiSlop Audit Report

**Slop Score**: 🚨 85/100
**Detected Voice**: [General AI] vs [{{voice_dna.tone}}]

### Block 1: Audit Report / Rapport d'Audit

- **Primary Issues**: [Issue 1], [Issue 2]
- **Verdict**: "Authentic", "Suspicious", or "Generated".

### Block 2: Version A (Polish / Lissage)

[Text with cleaned typography and grammar]

### Block 3: Version B (Humanize / Authenticité)

[Radical rewrite injecting {{voice_dna.idioms}} and concrete details]

### Block 4: Sparring Questions / Questions de Challenge

1. [Question targeting vague claim 1]
2. [Question targeting vague claim 2]

Source Credibility Check (Wiki Integration)

When the audited content cites wiki sources (references like [source](wiki/...) or claims marked with confidence levels):

1. Check confidence levels: For each cited wiki article, read its frontmatter confidence field.

   • high -> OK
   • medium -> Flag: "Source a confiance moyenne -- verifier avant publication."
   • low -> Flag: "Source a faible confiance -- rechercher des sources supplementaires."
   • Missing/no wiki -> Flag: "Affirmation sans source wiki. Ajouter via /copywriter:apprendre."

2. Check staleness: Read the updated field in cited article frontmatter. If older than 90 days, flag: "Source datee de plus de 90 jours. Re-rechercher avec /copywriter:recherche."

3. Report: Add a "Credibilite des sources" section to the audit report with per-source confidence and freshness scores.