Writing and Content Without the Slop
After this you can put your own thesis, voice, and load-bearing facts into the prompt *before* the model writes, so the draft comes back distinctive instead of generic, and you stop spending your editing time injecting the specifics the model was never going to invent.
Ask an empty model for a paragraph about your product and you get furniture. "We help companies grow." "In today's fast-paced landscape, businesses need scalable solutions." It reads fine. It is also indistinguishable from what your competitor's model produces from the same empty prompt, because both are reaching for the same place. A language model predicts the most probable next token, and the most probable phrasing of any common idea is the statistical average of everything written about it. Generic output is not a quality bug you prompt away with a better adjective. It is the default behavior of the machine, and the default is the center of the distribution.
This is why "make it sound more human" or "be more engaging" rarely moves the needle. Those instructions ask the model to be less average without giving it anything specific to be average about. The thing that pulls a draft off the center is information the model does not have and cannot guess: your actual numbers, the take your team argued about last quarter, the objection customers keep raising that the industry hasn't named, the words a real buyer used in a support ticket. A natural experiment makes the mechanism visible. When Italy temporarily banned ChatGPT in April 2023, Milanese restaurants' social posts became measurably more varied in vocabulary and structure, and that month saw roughly a 3.5% rise in average likes despite shorter and less frequent posting. When the tool came back, the homogenization came back with it. Distinctiveness is not extra effort layered on top. It is what is left when you stop routing through the averaging machine.
The operator move, then, is to do the origination work first. You supply the thesis, the proprietary facts, the voice, and an explicit do-not list, and then the model writes against that. This is not a hack; it is the difference in kind between the two ways people actually use these tools.
The deepest version of this fork is not about prompting at all. It is about which of two jobs you are actually doing. One camp treats the model as a volume machine: more posts, more variants, cheaper and faster, with the human writing a prompt and harvesting output. The other treats it as leverage for someone who already has judgment: the human supplies the thesis, the taste, the facts, and the accountability, and the model accelerates the mechanical middle. The entire quality gap between forgettable and credible lives in which camp you are really in, regardless of which one you would claim. Simon Willison's definition of slop is exactly this — content published without review and without anyone taking responsibility for it. Whether AI was involved was never the test; whether a named person would stand behind every claim in the piece is.
Where it breaks
The first failure mode is the one that gets you penalized: treating the abstraction default as a thing the editor cleans up afterward. It cannot be cleaned up afterward, because the defect passes a skim. Plausible, empty prose reads as competent. You scan it, nothing is wrong, you ship it. Editing AI output is a harder skill than writing fresh precisely for this reason — practitioners shipping at volume report that roughly 40% of AI prose needs significant revision, and teams that assume AI saves edit time ship the wordiness and end up slower net. The editor's real job is not polishing flow; it is injecting the situation-specific detail the model structurally could not know. If you did the origination work upstream, that injection already happened and the edit pass is light. If you didn't, no amount of editing rescues it, because you are editing furniture into more furniture.
The second is chasing the surface tells. The banned-word lists are real and the tells are real — delve, tapestry, seamless, the em-dash habit, the "it's not X, it's Y" frame, every paragraph closing on a tidy takeaway. But these are symptoms, not the disease. Find-replacing delve hides the smell while the hollowness remains, and paraphrasing is worse than useless: it keeps the same paragraph skeleton and swaps synonyms, so every cadence tell survives intact. The fix for hollow structure is to rewrite from the source, not to launder the words. The institutional version of this same error is running output through an AI detector and rewriting until it "passes" — that optimizes for evading detection, not for being good, and the detectors are unreliable in both directions anyway.
The third is the silent one, and it is the most dangerous because it has no alarm: voice drift, over distance and over time. Inside a single long piece, the model loses the early instructions as the context fills, and a voice you set at the top quietly decays by the bottom. Across many drafts, the edit pass gets lazier — aggressive on the first ten, permissive by the hundredth — until the output is AI-shaped wearing the writer's name, and the audience pattern-matches that drift within months. There is no gauge for either. You read it from symptoms, and the writer is the worst judge of their own drift. This is why voice has to be injected upstream and gated, not assessed by vibe at the end. You cannot edit voice into existence after the fact; you can only prevent its absence before the model starts.
The slop is in the empty prompt, so the fix is a block you assemble before the model writes a word. Open a fresh chat, fill these in with real material, and only then ask for the draft.
THESIS (the one non-obvious thing this piece argues — not a topic, a claim):
<e.g. "Repurposing kills voice at the variation step, not the drafting step.">
PROPRIETARY FACTS (numbers, names, dates the model cannot guess — at least one per section):
- <e.g. "Cut slop-pattern catches from 14/draft to 3/draft over 6 months by feeding corrections back into the brief.">
- <pull these from your own data, customer language, support tickets — verbatim, not paraphrased>
VOICE — write like THIS (3-5 sentences of my actual best writing, pasted verbatim):
<paste real samples; do not describe the voice, show it>
DO-NOT LIST (falsifiable bans, not adjectives):
- Never open with a rhetorical question.
- Never use "in today's fast-paced world" or "landscape."
- Never close a paragraph on a summary takeaway.
- <add the tells you personally hate>
TASK: Draft <the piece>, using the facts above. Every claim must trace to a fact I gave you;
if you need a number I didn't supply, write [NEED: <what>] instead of inventing one.Two lines carry most of the load. The do-not list is falsifiable where adjectives are not — "bold, witty, human" is unverifiable and the model ignores it, but "never open with a rhetorical question" is a rule a machine can actually obey or break. And the trace-or-flag line on TASK turns the model's instinct to fill gaps with plausible invented numbers into a visible [NEED:] marker you can resolve, instead of a fabricated statistic that survives your edit because it looks exactly like a real one.
One discipline makes the rest hold over time: when you catch a slop pattern in editing, add it to the do-not list and the voice block before the next draft, rather than just fixing it in place. A checklist catches errors one at a time forever. A brief that absorbs every correction reduces them — operators who feed fixes back report first-draft slop falling from low-teens patterns per draft to a handful over a few months. The brief is the durable asset; the prompt is disposable.
Worked example
IllustrativeIllustrative. A constructed before/after to show the move, not a real campaign.
A founder needs an "about" paragraph for a B2B landing page. The volume move is to ask and ship:
You: Write a short paragraph about what our company does. We help B2B SaaS companies with customer acquisition.
Assistant: In today's competitive landscape, acquiring customers is harder than ever. We partner with B2B SaaS companies to unlock scalable growth, streamline their acquisition funnel, and deliver measurable results that drive real impact. Our data-driven approach helps you reach the right customers at the right time.
Every sentence is true and every sentence is furniture. "Scalable growth," "measurable results," "real impact" — none of it could be falsified, and none of it is anything a competitor couldn't paste under their own logo. This is the abstraction default working exactly as designed: the model reached for the center and returned it. Now the same request, with one proprietary fact and a do-not list supplied first:
You: Draft an "about" paragraph. FACT: we reduced client acquisition cost by 34% across 12 accounts last quarter, mostly by killing paid channels that looked busy but converted no one. THESIS: most acquisition spend is hidden waste, not insufficient budget. DO-NOT: no "landscape," no "scalable," no rhetorical questions, no closing takeaway.
Assistant: Most acquisition budgets aren't too small — they're full of channels that look busy and convert no one. Last quarter we cut client acquisition cost by 34% across 12 accounts, mostly by turning off the spend that was buying impressions instead of customers. We start by finding the waste you're already paying for.
Same model, same length, a few seconds apart. The second one works not because it was prompted more cleverly but because it was given something true and specific to be about. The number is the spine; the thesis tells the model what the number means; the do-not list fences it away from the generic attractors it would otherwise drift toward. The author still owns the claim — that 34% had better be real and defensible, because the whole point of escaping furniture is that someone can now check you. That accountability is the feature, not a risk.