Stop Slop

Eliminate predictable AI writing patterns from prose.

Core Rules

1. Cut filler phrases and AI-marker words. Remove throat-clearing openers and emphasis crutches; replace excess LLM vocabulary (delve, underscore, foster, pivotal). See references/phrases.md.

2. Break formulaic structures. Avoid binary contrasts, dramatic fragmentation, rhetorical setups. See references/structures.md.

3. Vary rhythm. Mix sentence lengths. Two items beat three. End paragraphs differently.

4. Trust readers. State facts directly. Skip softening, justification, hand-holding.

5. Cut quotables. If it sounds like a pull-quote, rewrite it.

6. No em-dashes. Replace every em-dash (—), whether before a reveal, around an aside, or anywhere else, with a colon (introducing a list or expansion), commas or parentheses (around an aside), or a full stop. The em-dash is one of the strongest AI tells; the target in prose is zero. Keep dashes only in numeric ranges (8–25, an en-dash) and inside code.

Quick Checks

Before delivering prose:

• Three consecutive sentences match length? Break one.
• Paragraph ends with punchy one-liner? Vary it.
• Any em-dash (—) at all? Replace it with a colon, commas, parentheses, or a full stop. Zero em-dashes in prose.
• Explaining a metaphor? Trust it to land.

Two-Pass Audit

Before scoring, run one reflective pass:

1. Ask: "What in this draft makes it obviously AI-written?" Name the specific tells.
2. Fix exactly those. Don't rewrite blind — target what the first pass found.

This catches patterns the rules miss. The rubric below measures; this pass thinks.

Scoring

Rate 1-10 on each dimension:

Dimension	Question
Directness	Statements or announcements?
Rhythm	Varied or metronomic?
Trust	Respects reader intelligence?
Authenticity	Sounds human?
Density	Anything cuttable?

Below 35/50: revise.

Audit Mode

Use this when reviewing a draft cold, from a separate session or subagent that did not write it. Report only. Do not edit. The cold read is the point; rewriting throws it away.

Output, in order:

1. Tells found: a table, one row per concrete instance:

   Tell	Location (line / quote)	Fix

   Scan for em-dashes explicitly and flag every em-dash (—) in prose as its own row; the target is zero.

2. Rubric score: the five dimensions above, each 1-10, plus the total.

3. Verdict: pass (≥35/50, no Tier-1 LLM vocabulary, and no em-dashes in prose) or revise.

Quoting the exact line keeps the fixer honest. The writer applies the fixes; the auditor never touches the prose.

Examples

See references/examples.md for before/after transformations.

License

MIT