Remove AI Slop

Rewrite text so it no longer reads as machine-generated, while preserving the author's meaning, register, and voice.

Three principles

1. Preserve meaning. No new claims, no dropped facts, no editorialising.
2. Preserve register. Formal stays formal (just less floridly so). Casual stays casual.
3. Cut, don't rewrite, when you can. A lot of slop is padding. Deleting "It is important to note that" beats rephrasing it.

When to act, when to leave alone

Act when, in the same paragraph, two or more tells of any category co-occur OR three or more vocabulary-table words appear — that's the threshold for forced rewrite even if individually each one looks defensible. Single isolated tells are usually fine. (This is the unified rule in ../../shared/spec.md.)

Don't act on:

• A single em-dash in a paragraph
• One use of robust or leverage when it's the right word
• Lists when the content is genuinely list-shaped
• Formal register when context demands it (legal, academic, regulatory)
• Long sentences — long sentences are fine; slop is empty sentences

If the input is already clean, return it unchanged and add a single line: *No changes — text reads as written by a human.*

Categories of tells

Each category has a dedicated reference file. Read the reference only when you need it — when you've spotted that category of tell in the input and need the full table or fix recipe.

• Overused vocabulary — references/overused-vocabulary.md — the big word→alternative table
• Formulaic constructions — references/formulaic-constructions.md — "X, not Y", "not only X but also Y", copula avoidance, rule-of-three filler
• Padding and filler — references/padding-and-filler.md — promotional padding, hedging filler, superficial-analysis -ing tails, vague attribution
• Formatting tells — references/formatting-tells.md — em-dash overuse, bullet-and-bold soup, title-cased headings, excessive bolding, redundant summary sections
• Scaffolding and recap — references/scaffolding-and-recap.md — "Certainly!", "I hope this helps!", recap intros, "In conclusion" outros, "Despite challenges..." closers
• Elegant variation — references/elegant-variation.md — synonym-swapping that humans don't do

Apply user and project overrides

Before flagging vocabulary, read whichever of these exist (skip silently if absent):

1. ~/.claude/config/remove-ai-slop/user-vocabulary.md — the user's personal additions
2. ~/.claude/config/remove-ai-slop/user-allow-list.md — words the user never wants flagged
3. .claude/skills/remove-ai-slop/overrides/user-vocabulary.md — project additions
4. .claude/skills/remove-ai-slop/overrides/user-allow-list.md — project allow-list

Merge order: defaults → personal → project. Allow-list wins everything: if a word appears on any allow-list, never flag it, even if it's in the defaults or in another addition.

See overrides/README.md for the file format.

Output rules

Return the rewritten text only, in the same format as the input (plain text, markdown, etc.). No preamble. No "Here is the rewritten version:". No bullet list of changes.

Explanation is opt-in. Only explain what changed if asked ("what did you change", "show me the diff", "annotate"). If asked, follow with a short bulleted list under a ## Changes heading — one bullet per category of change, not one per edit.

Heavy-rewrite warning. If the input was so saturated with slop that the rewrite is materially different, end with: *Heavy rewrite — original meaning preserved but most sentences restructured.*

When given a file

Preserve all non-prose elements unchanged: code blocks, frontmatter, tables, image references, links. Only touch the prose. If prose is intertwined with structural markdown, edit in place and keep the structure.

Edge cases

• User wrote it themselves and is asking for a check. Same workflow — return the rewrite. Don't apologise or hedge ("This may have been written by a human, but…"). They asked.
• User wants only specific fixes ("just remove the em-dashes"). Do only that. Don't expand scope.
• Very short input (a single sentence). Apply the principles. If there's nothing to fix, say so in one line — don't invent problems.
• Translation or transcription. Some tells (formal register, slight stiffness) are translation artefacts, not AI tells. Be conservative.
• Text is intentionally in an AI voice (parody, satire, a quote of AI output). Ask before editing.

Examples

See examples/before-after.md for worked rewrites covering each category.