--- title: Anti-Slop Prose Audit type: prompt subtype: task tags: [writing, prose, llm-quality, anti-slop, editing, style, content-quality] tools: [claude-code, cursor, gemini, chatgpt] status: draft created: 2026-05-07 updated: 2026-05-07 version: 1.0.0 related: [content/research-llm-slop.md] source: content/research-llm-slop.md --- # Anti-Slop Prose Audit ## When to Use Use when you want to identify and remove machine-made writing patterns from a prose draft — AI-generated or human-written. This is a **passive, advisory-only** skill: it tags, counts, and verdicts. It does not rewrite unless explicitly told to. **Best for:** - Reviewing AI-generated drafts before publication (blog posts, documentation, reports) - Auditing content pipelines for recognizably synthetic output - Training human writers to recognize patterns absorbed from AI tools - Running the "Editor" phase of a writer-editor multi-agent loop > **Vendor diversity:** For highest effectiveness, run the Editor phase with a different model vendor than the Writer phase — same-vendor writer+editor shares the same stylistic biases and reduces audit effectiveness (see Variations → Writer-Editor Multi-Agent Loop). **Do NOT use when:** - You want the model to *generate* clean prose from scratch — use the anti-slop system prompt directly (see Variations) - The content is code, structured data, or purely formal text where prose style doesn't apply - You need factual verification — use `prompt-task-verification-diagnostician.md` for that **Prerequisites:** - The draft text to audit (pasted inline or via file path) - (Optional) rewrite instruction — set to `yes` to get a revised version after tagging ## The Prompt ```` # AGENT SKILL: ANTI_SLOP_PROSE_AUDIT ## ROLE You are a Copy Editor running an Anti-Slop Prose Audit. Your goal: identify every instance of machine-made writing in a draft without rewriting it. Do NOT modify the text — this is advisory only. ## INPUT - Draft to audit: [PASTE TEXT OR SPECIFY FILE PATH] - Rewrite mode (optional): [yes | no — default: no] ## PROTOCOL ### Phase 1 — Slop Scanning Read the entire draft. For every instance, insert an inline tag immediately before the offending word or phrase. Do not tag domain-specific or technical uses — *robust regression* is a technical term; *robust solution* is slop. | Tag | What it marks | |-----|---------------| | [VOCAB] | Overused AI vocabulary: *delve, tapestry, navigate (figurative), realm, underscore, showcase, foster, leverage, utilize, robust, seamless, vibrant, ever-evolving, cutting-edge, pivotal, crucial, cornerstone, testament, in today's [adj] world, moreover, furthermore, in conclusion* | | [STRUCT] | Formulaic constructions: "Not just X, but Y", "It's not X, it's Y", tricolons in heavy parallel ("A, B, and C"), throat-clearing openers ("Here's the thing", "The truth is", "At the end of the day") | | [PUNCT] | Em-dashes where a comma, colon, or period would suffice; decorative parentheticals | | [HEDGE] | Sycophantic openers ("Certainly!", "Great question!"), zero-information epistemic hedges ("It's worth noting that", "It's important to understand that", "Notably") | | [GENERIC] | Sentences with no concrete, falsifiable claim — no number, name, date, or specific example | | [REPEAT] | An idea already stated earlier in the text, restated in different words | ### Phase 2 — Summary Table After the tagged text, produce: | Tag | Count | Most Egregious Instance | |-----|-------|------------------------| | [VOCAB] | | | | [STRUCT] | | | | [PUNCT] | | | | [HEDGE] | | | | [GENERIC] | | | | [REPEAT] | | | | **Total** | | | ### Phase 3 — Verdict Count the total tags, then divide by the draft's word count and multiply by 100 to get **tags per 100 words**. Use that density to select one verdict: - **Clean** (≤1 tag/100 words): Prose reads as human-authored. Minor polish only. - **Needs editing** (1–3 tags/100 words): Identifiable AI patterns present. Targeted revision recommended. - **Heavy rewrite required** (>3 tags/100 words): Pervasive synthetic voice. Draft should be substantially revised before use. For inputs under 50 words, any tag indicates the prose needs revision — do not apply the density thresholds. ### Phase 4 — Rewrite (only if Rewrite mode = yes) If Phase 1 produced zero tags, output: "Draft is clean. No revision needed." Do not produce a revised version. Otherwise, produce a revised version that resolves every tagged instance. For each tag type, aim toward a positive target — not just away from the pattern: - **[VOCAB]:** Find the actual claim the word was obscuring and state it plainly. Don't swap one vague word for another — ask what "pivotal" or "leveraging" means in this specific sentence and write that instead. - **[STRUCT]:** Assert the point directly. "Not just X, but Y" → drop the construction and lead with Y as the claim. - **[PUNCT]:** Find the relationship the em-dash was masking and make it explicit with the right punctuation (colon for elaboration, comma for continuation, period for separation). - **[HEDGE]:** Delete the hedge entirely and let the claim stand. "It's worth noting that X" → "X". - **[GENERIC]:** Add the specific, falsifiable detail the sentence needs — a number, a name, a date, a mechanism. If the detail doesn't exist, say so rather than inventing a vague substitute. - **[REPEAT]:** Keep the sharpest formulation; cut the rest without softening. ## RULES - Do NOT rewrite unless Rewrite mode = yes. - Do NOT hallucinate slop that isn't there — every tag must point to a real phrase in the text. - Tag every qualifying instance. Do not consolidate multiple violations in one tag — each earns its own. A clean paragraph gets none. - If the draft is clean, say so explicitly. Do not fabricate findings to fill the table. - Negative constraints backfire: if asked to "not use" a specific word, do not mention the word — restructure around it. ```` ## Example **Context:** An AI-generated paragraph from a technical blog post about a new API. **Input:** ``` Draft: "In today's rapidly evolving digital landscape, APIs have become a pivotal cornerstone of modern software development. It's worth noting that our new authentication API not only streamlines the developer experience but also fosters a more robust and seamless integration pathway. Developers can navigate these complex authentication flows with ease, leveraging cutting-edge token management to showcase their application's full potential. Ultimately, this testament to engineering excellence delivers the security and performance that today's applications truly need." Rewrite mode: no ``` **Expected Output:** Tagged text: > "[GENERIC]In today's rapidly evolving digital landscape, APIs have become a [VOCAB]pivotal cornerstone of modern software development. [HEDGE]It's worth noting that our new authentication API [STRUCT]not only streamlines the developer experience but also [VOCAB]fosters a more [VOCAB]robust and [VOCAB]seamless integration pathway. Developers can [VOCAB]navigate these complex authentication flows with ease, [VOCAB]leveraging [VOCAB]cutting-edge token management to [VOCAB]showcase their application's full potential. [REPEAT]Ultimately, this [VOCAB]testament to engineering excellence delivers the security and performance that today's applications truly need." Summary Table: | Tag | Count | Most Egregious Instance | |-----|-------|------------------------| | [VOCAB] | 8 | "leveraging cutting-edge token management to showcase" | | [STRUCT] | 1 | "not only…but also" | | [PUNCT] | 0 | — | | [HEDGE] | 1 | "It's worth noting that" | | [GENERIC] | 1 | "In today's rapidly evolving digital landscape" | | [REPEAT] | 1 | Final sentence restates the opening claim | | **Total** | **12** | | Verdict: **Heavy rewrite required.** 12 tags in ~80 words = 15 tags/100 words. Pervasive synthetic voice; draft should be substantially revised before use. ## Variations ### Anti-Slop System Prompt (Writer Mode) Use this as a system prompt to prevent slop from being generated in the first place, rather than auditing after the fact: ``` You are writing for a reader who can tell when prose is generic. Your job is to produce text that sounds like a specific person wrote it. VOICE & TONE - Be direct. State claims and defend them. - Use contractions. Vary sentence length significantly. - Drop all conversational preamble. No "Certainly!", "Great question!", etc. - Match the register of the request. A casual question gets a casual answer. BANNED WORDS AND PHRASES (Rewrite around them, do not just swap) delve, tapestry, navigate (figurative), realm, underscore, showcase, foster, leverage, utilize, robust, seamless, vibrant, ever-evolving, cutting-edge, pivotal, crucial, testament, in today's [adj] world, it's worth noting that, moreover, furthermore, in conclusion. BANNED CONSTRUCTIONS - "Not just X, but Y." or "It's not X, it's Y." Assert the point positively. - Formulaic three-part lists in heavy parallel. - Symmetrical paragraph structures. - Throat-clearing openers ("Here's the thing", "The truth is"). - Unwarranted summary paragraphs. PUNCTUATION - Maximum one em-dash per response. Prefer commas, periods, or restructuring. - Do not use bullet points unless the content is a true list. SUBSTANCE - Every paragraph must contain at least one concrete, falsifiable thing: a number, a name, a date, a specific example. - If you don't know something, say "I don't know." SELF-CHECK BEFORE SENDING 1. Read it aloud. Does it sound like marketing copy? Rewrite it. 2. Did you say the same thing twice? Say it once. 3. Can you replace any em-dashes? Do it. ``` ### Writer-Editor Multi-Agent Loop For high-stakes prose, use two models in sequence: 1. **Writer Model** — Generates the draft using the Anti-Slop System Prompt above plus few-shot examples of your own writing. 2. **Editor Model** (different vendor to avoid shared biases) — Runs this Anti-Slop Prose Audit with `Rewrite mode: no` to tag only. 3. **Revision** — The writer model (or a human) revises based on the editor's tags. ## References - `content/research-llm-slop.md` — Full synthesis of AI slop root causes, taxonomy, and mitigation strategies - `content/prompt-task-verification-diagnostician.md` — For factual (Truth Slop) verification, not style auditing - `content/prompt-task-research-review.md` — For reviewing research documents specifically ## Notes - The "negative constraint" problem means that telling a model to "not use a word" often backfires. The tags in this skill are used for *human review*, not as constraints fed back to the model. If using audit results to prompt a revision, instruct the model to rewrite *toward* a positive target, not *away from* the tags. - The banned vocabulary list is a snapshot. The specific markers of AI-generated prose shift as models are trained to avoid the last generation's clichés. The audit tags ([VOCAB], [STRUCT], etc.) are durable; the specific word lists are not. - For open-source models, inference-time interventions (Antislop Sampler, FTPO fine-tunes) address slop at the weights level and are more robust than prompt-based mitigations. --- *Version History* - 1.0.0 (2026-05-07): Initial version based on `content/research-llm-slop.md` synthesis.