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.

1. Extract every heading, in order, with no body text. Read that list alone. It must read as a rising argument, not a table of contents — a flat topic list means the piece is organized by subject, not by claim.
2. Check the bookend. Does the piece open with the question or tension it answers, and does it close by resolving that same tension — not by restating the opening in different words? A piece that stops after the last data point, with no opening frame and no landing, is structural slop even when every sentence individually passes the prose pass.

Code pass (diff-scoped)

1. Scope is the diff — branch vs main, or the files just written. Do not expand into general cleanup uninvited.
2. Hunt, in order: comments that narrate code or restate names · defensive checks and try/catch on trusted internal paths · any / @ts-ignore / # type: ignore used to bypass types · deep nesting fixable by early returns · needless abstraction and generic names (handleData, processItem, result) · test slop (ask: would this test fail if the function broke?).
3. Two-way defensive check: real defenses stay — timeouts, retries, rate limits on external calls are not slop. Removing all error handling is overcorrection.
4. Behavior unchanged unless fixing a clear bug. One smell per pass, smallest diff, match surrounding style. Summary in 1–3 sentences.

Hard rules

• Never invent specifics. Names, numbers, dates, quotes, citations come only from the source text, the conversation, or verification you actually ran. Missing detail → leave a [ADD: which study?]-style placeholder, or use a simpler sentence.
• Em dash is budgeted, not banned: at most one per ~150 words, never the paired-aside form — like this —, never two in a sentence.
• Preserve force-bearing words exactly — "never", "must", "all" in safety, security, legal, and technical rules do not get softened.
• No manufactured humanity: no fake typos, no forced slang, no persona. Humanizer-tool output is its own genre of slop.
• Final audit before returning: ask "what still makes this obviously AI-generated?" — fix it or name it in the report.