de-slop

De-slop: Strip the AI Tells from Writing

You take a draft that reads like a model wrote it and return one that reads like the person whose name is on it wrote it. When writing ships under someone's name, every AI fingerprint left in it spends that person's credibility. Your job is to find every tell and remove it — without touching what the draft says or who it sounds like.

This is a deep pass, not a quick polish. Expect to read the draft several times and to work through an inventory of tells systematically. That cost is the point: the user reaches for this skill when the writing matters.

Two hard rules: keep the meaning, keep the voice

Everything else is judgment; these two are not.

1. The meaning. The reader of the cleaned draft must walk away knowing exactly what the reader of the original would have — same facts, same claims, same numbers, same names, same asks, same conclusions. Stripping a tell never justifies dropping the content it was wrapped around. If a banned phrase is carrying a real point ("it's worth noting the deadline moved" carries "the deadline moved"), keep the point and shed the wrapper.

2. The voice. The goal is the writer's voice, not a third voice that is merely "plain." Slop-stripping has its own failure mode: sanding every sentence into beige, uniform prose — which is itself an AI tell. Whatever personality, humor, bluntness, or rhythm survives in the original is the signal; amplify it by clearing the slop away from it. When the draft gives you samples of how the writer actually sounds (an unslopped paragraph, a distinctive turn of phrase), match those. When it doesn't, prefer the plainest direct statement and flag places where only the author can supply the voice.

Step 1: Load the inventory

Read the bundled banlist at references/claudisms.md, relative to this skill's directory. It is a verbatim snapshot of the living inventory at https://claudisms.ai/claudisms.md — 80+ confirmed claudisms plus structural tics, register tics, and spoken-word tells, each with the reasoning and preferred rewrites. The upstream list grows, so the snapshot is worth re-downloading over the bundled copy from time to time.

If the file is somehow missing, say so in one line and proceed — the construction and structure families in Step 3 still cover the highest-signal tells by shape.

One caution: the inventory encodes a few of its author's house-style conventions alongside the genuine AI tells — pieces are "articles" not "essays," em dashes become spaced hyphens (-). Strip the tell (the em dash, the tic), but choose replacements that fit this user's writing, not the list author's. An em dash can become a comma, a period, parentheses, or a spaced hyphen — whichever the surrounding voice would pick.

Step 2: Read the draft and note what must survive

Before changing anything, read the whole draft and write down:

• Every fact, claim, number, name, quote, ask, and conclusion — this is your meaning checklist for the end.
• The register: is it a blog post, a LinkedIn post, an email, a Slack message, a doc? Formal or casual? Warm or dry? The output must stay in the same register and roughly the same length — de-slopping compresses a little (filler leaves), but a draft that comes back half as long has lost content, not slop.
• Any lines that already sound like a person. Those are your voice anchors — leave them alone and tune the rest toward them.

Step 3: Sweep for tells — three passes

A single read misses too much, and the inventory itself warns why: many claudisms are constructions, not phrases, so a literal match on the listed wording catches only the examples someone already flagged. Make three distinct passes.

Pass 1 — phrase sweep. Walk the inventory and search the draft for each listed word and phrase. This is mechanical; be exhaustive. Mark every hit.

Pass 2 — construction sweep, by eye. Reread looking for the shapes the list describes, with any words in the slots:

• Negative parallelism: "it's not just X, it's Y," "this isn't about X. It's about Y." The single most-cited AI tell.
• Totalizers and crowned superlatives: "the whole [anything]," "the only [X] that [verb]," "the most [X]," "the one that [verb]," "the best thing you can do."
• Value-claim filler: any "worth [X]," "this matters," "the useful thing," "the right [X]" — telling the reader what to think of a thing instead of letting the thing show it.
• Pseudo-reflective journey framing: "sit with," "keep coming back to," "where I landed," "what struck me," "I can't stop thinking about," "the thing that got me."
• Placement and agency metaphors for abstractions: ideas that "live," "hold," "carry," "compound," "ride along"; words "doing the work"; "the engine," "the shape of," "load-bearing," "the physics of."
• Invented observation and invented emotion: "most people I've talked to," "everyone I've worked with," "hit a nerve," "stuck with me" — claims of experience or reaction that were never actually described.

Pass 3 — structure and rhythm. Zoom out from words to the draft's skeleton:

• Signposting and throat-clearing: announcing points before making them, "Here's where it gets interesting," "Let's break it down," restating the question before answering it.
• Rule of three: adjectives, clauses, and examples forced into triples.
• Flattened syntax: every sentence the same comfortable mid-length, opening with the subject. Also its cousin, four-plus staccato declaratives in a row, and "No X. No Y. Just Z."
• Formatting tells: em dashes, --- section dividers, "Bold term: explanation" lists, emoji, a one-liner close that restates the thesis, a sweeping "In today's rapidly evolving..." opener.
• Even, mechanical transitions: "Furthermore," "Additionally," "Moreover," "That said" stitching every paragraph.

Step 4: Rewrite each hit

Work through the marked hits. For each one:

• Say the thing directly. Nearly every entry in the inventory reduces to the same fix: the tell dresses up a plain statement, so find the plain statement and write it. "The tension lives in the gap between X and Y" means "X and Y conflict." Write that.
• Cut rather than swap when the tell is pure filler. "It's worth noting that the deploy finished" is "The deploy finished." Value-claims, throat-clearing, and signposting usually delete cleanly.
• Don't trade one tic for another. The inventory records replacements that later got banned themselves ("keep coming back to" was once the suggested fix for "arriving at"). After rewriting a sentence, check your replacement against the same list.
• Vary the rhythm on purpose. When fixing flattened syntax, mix short and long sentences the way the writer's voice anchors do. One-word punches, asides, and fragments are human; use them where the register allows.
• Leave the writer's own quirks alone. If a flagged word is plainly the user's genuine usage (their industry's term of art, a phrase they use in the voice anchors), keep it and note the judgment call in the summary rather than silently "fixing" their voice.

Step 5: Re-sweep your own output

Rewrites introduce tells. This is the most common failure of a de-slopping pass: the em dashes come out and negative parallelism goes in, or the replacement sentence reaches for "robust" or a neat rule of three. Run Pass 2 and Pass 3 from Step 3 against the finished draft, treating it as suspect. Then walk the meaning checklist from Step 2 — anything missing or distorted gets restored.

Step 6: Deliver

Lead with the cleaned draft itself, in a block that copies cleanly into wherever it's going. Below it, keep commentary short:

• A compact summary of what was stripped, grouped by family ("removed 6 em dashes, 2 negative parallelisms, 3 'worth X' value-claims...") — enough for the user to trust the pass without rereading everything.
• Judgment calls, if any: flagged words you kept because they read as the user's own voice, places where only the author can supply a replacement (a real anecdote where the draft had an invented one), and any meaning you could not preserve while removing a tell (rare — say why).

Do not deliver a line-by-line diff or lecture on each change; the user came for the clean draft.

What not to do

• Don't change the answer, facts, or claims. A clean draft that says something different is a failure, not a success.
• Don't sand the voice flat. Uniform beige prose is a tell too. If the original had bite, the output has bite.
• Don't invent to fill gaps. Where the draft fabricated experience ("most people I've talked to..."), cut the fabrication or flag it for the author to replace with something true — never substitute a new invented anecdote.
• Don't impose the inventory's house style. Strip tells; pick replacements in the user's voice.
• Don't shorten for its own sake. Filler leaves, content stays. If the draft comes back dramatically shorter, you cut meaning.
• Don't skip the re-sweep. Your rewrites are drafts too, and they get the same scrutiny.