aio-anti-slop

Anti-slop

Overview

Slop is the absence of a decision: the statistically likely filler nobody chose. Two failure modes both count as slop — AI tells left in, and overcorrected "anti-slop voice" (staccato fragments, forced slang, fake casualness). You cannot reliably see your own slop; the priors that produce a pattern make it invisible on re-read. So: detect mechanically, judge each hit in context, repair minimally.

This is a style pass, not a content edit. Meaning, claims, facts, and code behavior stay identical.

Prose pass

1. Anchor genre and voice. If the target genre (docs, blog, tweet, academic) or the author's voice is unclear, ask one question first. The author's existing voice overrides every rule below.
2. Scan mechanically using the catalog at references/patterns.md (relative to this SKILL.md) — run the rg commands, then do the by-hand checks. A match is a lead, not a verdict.
3. Triage every hit: defect or protected use. Protected: quotations, code blocks, proper names, domain terms used precisely, and the author's own recognizable habits. Edit only when the defect is clearer than preservation — when unsure, no-op. A no-op is a first-class outcome.
4. Repair finding by finding. Decide what the sentence actually asserts, then assert that. Never fix a pattern by paraphrasing the pattern — the catalog lists banned escape hatches that count as new findings. Sentences without findings are copied byte-for-byte.
5. Re-scan after rewriting (rewrites reintroduce slop). Maximum 3 passes; a pattern surviving 3 passes gets rewritten as a bare claim, not paraphrased again.
6. Report: hits found → fixed → intentionally kept (with the reason), per pattern.

Structural pass (reports, explainers, write-ups only)

Skip this pass for anything without a throughline — tweets, changelogs, reference docs, single paragraphs. Word-level slop and structural slop are independent: a piece can pass every check above sentence by sentence and still read as a raw dump if it never opens or never lands.