eliminating-ai-slop

Eliminating AI Slop

Guidelines: See CLAUDE.md for writing standards. Last Updated: 2026-03-13 See also: reference.md (patterns), examples.md (usage)

Wrong skill? Analyzing/scoring text (read-only) → detecting-ai-slop. Profanity/inappropriate language → professional-language-audit.

Scope

Fires for: All human-readable prose — messaging, email, social/professional, documentation, business writing. Does NOT fire for: AI-to-AI content (prompts, system instructions, agent config, tool parameters, few-shot examples).

Two Modes

1. Interactive — User provides text → show flagged patterns with suggestions → user picks: "Rephrase all", "Keep all", "List them", or "Rephrase 1,3,5"
2. Automatic (GVR Loop) — Generate → Verify (check patterns + stylometrics) → Refine (max 3 iterations). Report: [GVR: 2 iterations | removed 8 patterns | σ: 7.2→16.4]

GVR Thresholds

Metric	Pass	Action if Failed
Lexical patterns	0	Rewrite flagged phrases
Sentence length σ	≥15.0	Vary sentence lengths
TTR	0.50-0.70	Diversify vocabulary

Rewriting Guidelines

1. Preserve meaning — Change how, not what
2. Increase specificity — "incredibly powerful" → "handles 10K concurrent connections"
3. Vary structure — Break uniform patterns
4. Delete over replace — If a phrase adds nothing, cut it
5. Commit to positions — "It depends" → "Use X for <1000 users, Y for more"

Dictionary

Location: {workspace_root}/.slop-dictionary.json — this skill writes, detecting-ai-slop reads. Commands: "Add [phrase] to slop dictionary" | "Never flag [phrase]" | "Show my top slop patterns"

Quick-Reference: Common Patterns

Slop Pattern	Better Alternative
"It's worth noting that..."	(delete — just state it)
"In order to..."	"To..."
"Leveraging/utilizing"	"Using"
"A comprehensive solution"	(describe what it actually does)
"Incredibly powerful"	(specific metric or capability)
"Seamless integration"	"Connects to X via Y"
"It's important to understand"	(delete — just explain)

See reference.md for the full pattern catalog.

Self-Check

Before publishing: meaning preserved? specificity added? voice consistent? no new slop introduced? GVR thresholds met?

Companion Skills

detecting-ai-slop (analysis, read-only) | professional-language-audit (profanity detection)

• detecting-ai-slop: Read-only analysis (this skill is the active rewriter)
• readme-authoring: Slop prevention in READMEs
• incorporating-research: Clean up pasted research text
• markdown-table-discipline: Slop prevention in table content

When to Use

• When authoring any human-readable prose (docs, email, messages, tickets)
• When wiki-orchestrator pipeline triggers slop detection stage
• When reviewing AI-generated content before publishing

Failure Modes

Failure	Fix
Over-correction strips personality from writing	Preserve author voice — only target known slop patterns
False positive on legitimate hedging language	Context matters — "it's worth noting" in a risk section is fine
Slop patterns evolve faster than the deny list	Update pattern list quarterly from real examples

# Example: invoke slop detection
node ~/.codex/superpowers-augment/superpowers-augment.js use-skill eliminating-ai-slop