---
name: slop-check
description: 'Assess a piece of writing for AI slop — hollow, templated, machine-sounding prose that looks finished but says little. Use whenever the user asks to "check for slop," run a "slop check," whether something "sounds AI-written" or "sounds like AI," whether a draft is hollow, generic, padded, or templated, or to audit writing for AI tells. Works on pasted text, markdown, .docx, and .pptx files. Diagnostic only: it reports findings and a verdict, it does not rewrite. Do not use for code review or visual design review.'
---

# Slop Check

Slop is writing with superficial competence and nothing under it — fluent, polished, hollow. The tell is a gap: the text looks more finished than it is informative. There is an asymmetry of effort, too — it was cheap to produce and it is tiring to read, because each sentence costs the reader more than it returns.

This skill diagnoses that gap. It does not fix it. It reports what is wrong and where, and the user does the rewriting. Reporting without rewriting is deliberate: the user keeps their own voice, and a problem they fix themselves is a problem they stop making.

## Step 1 — Get the text

The skill assesses text. It may arrive several ways — handle whichever applies:

- **Pasted text, markdown, plain text** — use directly.
- **A local file path** — read it. For `.docx`, extract with the `docx` skill; for `.pptx`, extract slide text and speaker notes with the `pptx` skill.
- **A file in a connected source** (Google Drive, SharePoint, etc.) — fetch the content with that connector's tools, then assess it. If the source is not connected in the current environment, say so plainly and ask the user to paste the text or upload the file — do not pretend to reach it.
- **An uploaded document** (common on claude.ai) — the content is usually already extracted into context; use that rather than re-extracting.
- **A folder, or several files at once** — assess each artifact separately and produce one report per artifact. Do not merge them into a single verdict; a folder is not one piece of writing. If there are many files, list what you found and confirm scope with the user before running all of them.

File-type notes:
- **.pptx** — assess the text only. Visual design — layout, typography, color — is out of scope; say so once and move on.
- **.xlsx and spreadsheets** — out of scope. Slop is a prose failure, and a spreadsheet has little prose to assess. If a sheet has narrative cells or a notes tab the user cares about, ask them to paste that text in.

## Step 2 — Name the artifact type

State what you are assessing: memo, essay, email, report, slide copy, casual message. It changes the verdict. A construction that is slop in an investor memo is fine in a Slack message; bullet fragments are slop in an essay and correct in slide copy. If the type is unclear, infer it, say what you inferred, and judge against that.

## The core principle: hollowness is the disease, tells are symptoms

The five passes below split into two kinds. Passes 3–5 test whether the writing actually says anything — they find the disease. Passes 1–2 catch surface markers — they find symptoms.

Weight them accordingly. A piece can carry a few flagged words and still be sharp, earned writing. A piece can contain zero flagged words and be pure slop — hollow, hedged, pointless. Never flag a word for being on a list. Flag it when it sits on top of an empty sentence. The lexical pass corroborates; it does not convict.

## Step 3 — Run the five passes

### Pass 1 — Lexical tells (corroborating)

Read `references/ai-tells.md` for the catalog. It is tiered: Tier 1 words are flagged on sight, Tier 2 only when clustered, Tier 3 only at density. The tiering exists to kill false positives — "significant" is not slop; five "significant"s on a page is. Note clustered tells and template phrases, with locations.

### Pass 2 — Structural tells (corroborating)

Look for the shape of machine writing: uniform paragraph length, compulsive rule-of-three, the "not just X, but Y" negation flip, the hedging seesaw ("However… That said… On the other hand…"), bullet lists of bolded fragments, rhetorical questions as section openers, em-dash overuse, and a conclusion that restates the body without adding to it. See `references/ai-tells.md` for the full list.

### Pass 3 — The empty-sentence test (primary)

Take each sentence and delete it. Does the reader lose anything — a fact, a claim, a turn in the argument, a reason to keep reading? If nothing is lost, the sentence is slop. Throat-clearing ("Let's explore this further"), restatements of the prompt, and transitional filler all fail this test. The standard: every sentence should change what the reader knows or what they do next. Quote the worst offenders.

### Pass 4 — The specificity test (primary)

Find the claims. Then check whether each is anchored — to a number, a name, a date, a concrete example, a mechanism — or whether it floats. "This drives meaningful efficiency gains" floats. "This cut the close from nine days to four" lands. Confident vagueness is the most expensive kind of slop, because it reads as substance and is not. Flag claims that assert without anchoring.

### Pass 5 — Voice and register (primary)

Ask whether a specific person with a point of view wrote this, or whether it was assembled to give offense to no one. The committee tells: false balance where the writer will not take a side, neutral corporate register, hedging on every claim, no opinion the writer would defend under push-back. A real argument can be wrong; slop cannot be wrong, because it never commits to anything. Flag the absence of a position where the piece needs one.

## Step 4 — Output

Structure the report exactly like this.

### Verdict

One of three, with one sentence of justification:
- **Clean** — reads like a person wrote it on purpose. Nits at most.
- **Some slop** — recognizable tells; a focused pass would sharpen it. Not pervasive.
- **Heavy slop** — hollowness throughout. Looks finished, says little.

No numeric score. A "62/100" is false precision dressed as rigor.

### Lead finding

The single worst thing in the piece. Quote it, locate it, say why it fails. Lead here because if the user fixes only one thing, it should be this one.

### Findings

Ranked by severity, worst first. For each:
- **Quote** the offending text and give its location.
- **Pass** it failed.
- **Why** it is slop, in plain terms.
- **Fix direction** — where to take it. Not a rewrite. A direction: "cut this," "anchor this to a number," "take a side here."

Stop there. No "what's working" section, no encouragement, no closing summary that restates the findings. The verdict and the findings are the whole report.

## The standard you are judging against

Slop is the absence of these. Judge the writing against them:

- **Every sentence earns its place.** It changes what the reader knows or does. If it does not, it goes.
- **Concrete over abstract.** Numbers, names, examples, mechanisms. Abstraction with no path back to the concrete is where slop hides.
- **A position, held.** The writer commits to something falsifiable. Hedging everything to avoid being wrong is its own failure.
- **Plain words.** Orwell's rules: short word over long, cut any word you can cut, active over passive, no stale metaphor, no jargon where plain English exists.
- **No padding.** No preamble, no throat-clearing, no validating language, no conclusion that restates the body.
- **Form matches substance.** Writing that is analytically sound but structurally incoherent — uniform, templated, shapeless — still fails. The shape is part of the argument.