Diagnostic · symptom → fix

LLM Prose Smells

Like code smells, a prose smell is not a bug—the text is grammatical and often fluent. It is a surface symptom that usually points to a deeper problem: padding, hedging, sameness, sheen. This is a field guide to the tell-tale signs of AI slop, each paired with a copy-ready prompt that fights back.

How to use: Spot the symptom in a draft, grab the fix prompt, and send it. Each fix maps to one or more concepts in the reference.

Rule of thumb: Name the exact behavior you want gone. "Be less wordy" is weak. "Cut every sentence that does not add a fact; target 150 words" is a control.

The catalogue

1. Padding & Preamble

Symptom: Throat-clearing before the answer arrives.

Tells: "Certainly!", "Great question.", "In today's fast-paced world…", "Let's dive in."

SmellCertainly! That's a great question. Let's dive into the key considerations you'll want to keep in mind when choosing a database.

FixedChoose Postgres unless you have a specific reason not to.

Fix prompt

Answer with no preamble. Delete any opening pleasantry, restatement of my question, or "let's dive in" framing. Lead with the single most important sentence (BLUF), then support it. First word should carry information.

See BLUF · Minimalist Instruction

2. Runaway Length

Symptom: A three-line answer inflated to three paragraphs.

Tells: Restated points, "furthermore", "additionally", padding to feel thorough.

SmellThere are a number of factors to consider. Firstly, it is important to note that… Furthermore, one should also bear in mind… Additionally, it is worth mentioning…

FixedTwo factors matter: cost and latency.

Fix prompt

Respond in 150 words or fewer. Every sentence must add a new fact—cut any that only restate, transition, or reassure. If you cannot fit it, drop the least important point rather than compress everything into mush.

See Reading-Time Calibration · Hit a target reading time

3. Corporate Sheen

Symptom: Marketing gloss where plain words would do.

Tells: "leverage", "robust", "seamless", "delve", "cutting-edge", "unlock", "elevate", "in the realm of".

SmellOur robust, cutting-edge solution empowers teams to seamlessly leverage data and unlock actionable insights.

FixedTeams can query the data to make decisions.

Fix prompt

Rewrite in plain, standard English. Ban this vocabulary: leverage, robust, seamless, seamlessly, delve, unlock, elevate, empower, harness, cutting-edge, actionable, synergy, holistic, in the realm of, at the end of the day. Replace each with a concrete, ordinary word. No marketing tone.

See Basic English & Restricted Lexicons · Controlled Vocabulary

4. Hedge Haze

Symptom: So many qualifiers the claim dissolves.

Tells: "it's important to note", "generally", "in many cases", "may potentially", "could arguably".

SmellIt's worth noting that this could potentially, in many cases, generally lead to somewhat improved performance.

FixedThis improves performance by about 20%. It does not help when the cache is cold.

Fix prompt

State claims with commitment. Remove filler hedges (it's worth noting, generally, in many cases, potentially, arguably, somewhat). Keep a qualifier ONLY where the uncertainty is real and specific—then name the exact condition or number instead of a vague hedge.

See Hedging vs. Commitment

5. Sycophancy

Symptom: Flattery and validation padding the reply.

Tells: "Great question!", "You're absolutely right!", "What a fascinating idea."

SmellGreat question! You're absolutely right to think about this. That's a really smart approach.

FixedThat approach breaks under concurrent writes. Use a queue instead.

Fix prompt

Drop all praise, flattery, and validation. Do not tell me my question is good or my idea is smart. Do not agree by default—if I am wrong, say so and explain why. Prioritize accuracy over affirmation.

See Minimalist Instruction

6. The Summary Sandwich

Symptom: Restates the question up front, re-summarizes at the end—content only in the middle.

Tells: "To summarize…", "In conclusion…", "As we discussed above…" closing a short answer.

SmellYou asked how to reverse a list. Here's how to reverse a list… In summary, that is how you reverse a list.

FixedUse list.reverse() to reverse in place, or reversed(list) for an iterator.

Fix prompt

Do not restate my question and do not add a closing summary or "in conclusion" paragraph. For a short answer, give the answer once. Only include a summary if the response is long enough (roughly 500+ words) to actually need one.

See Inverted Pyramid · BLUF

7. Rule-of-Three Tic

Symptom: Everything arrives in triples, whether the idea has three parts or not.

Tells: "clear, concise, and compelling", relentless three-adjective and three-item rhythm.

SmellThis is a powerful, flexible, and intuitive tool that saves time, reduces errors, and boosts morale.

FixedThis tool is flexible and cuts input errors by half.

Fix prompt

Vary sentence structure. Do not default to sets of three (triads of adjectives or list items). Use the number of items the idea actually has—often one or two. Never stack three synonyms where one precise word works.

See Isocolon / Tricolon (use it on purpose, not by reflex)

8. Elegant Variation

Symptom: The same thing is renamed for "variety", so the reader wonders if it changed.

Tells: "the user… the account holder… the individual… the end-user" for one entity.

SmellThe user logs in. The account holder then picks a plan. Finally, the individual confirms.

FixedThe user logs in, picks a plan, then confirms.

Fix prompt

Use one term per concept throughout. Pick the single best name for each entity and repeat it—do not swap in synonyms for variety. Repetition of the key noun is correct here, not a flaw.

See Controlled Vocabulary · Anaphora & Reference Resolution

9. Nominalization Fog

Symptom: Verbs buried inside abstract nouns; actors disappear.

Tells: "the utilization of", "perform an analysis of", "make a determination", "-tion / -ment" pileups.

SmellThe implementation of the optimization resulted in a reduction of latency.

FixedWe optimized the query, so it runs faster.

Fix prompt

De-nominalize: turn abstract nouns back into verbs and name who does what. Replace "the X-tion of Y" with "someone X-es Y". Prefer strong verbs over "make / perform / conduct / carry out" plus a noun. Cut empty subjects like "there is / it is".

See De-nominalization · E-Prime

10. Listicle Overload

Symptom: Every answer is bullets and bold headers, even when a sentence would be clearer.

Tells: Nested bullets for two related ideas; a bulleted list where a flowing argument belongs.

Smell**Reasons:**
• It is fast
• It is cheap
• Because of these it scales

FixedIt is fast and cheap, so it scales.

Fix prompt

Write in prose paragraphs by default. Use a bulleted list only for genuinely parallel items (3+) or sequential steps. Do not bullet a single idea or a cause-and-effect that reads better as a sentence. No decorative bold on every line.

See Chunking & List Discipline

11. Vacant Abstraction

Symptom: General claims with no concrete example, number, or instance.

Tells: "various factors", "a range of solutions", "different approaches"—never a named one.

SmellThere are various strategies you can use to improve performance in different scenarios.

FixedAdd an index on user_id; it cut our query from 400ms to 12ms.

Fix prompt

Ground every general claim in a concrete specific: a named example, a number, or a real instance. Replace "various / several / a range of X" with the actual X. If you cannot name a specific, cut the sentence.

See Lexical Density · Minimalist Instruction

12. Formatting Confetti

Symptom: Emoji, bold, and headers scattered for decoration, not meaning.

Tells: 🚀 section headers, **bold** on random phrases, an emoji ending every bullet.

Smell🚀 **Getting Started** — Let's **supercharge** your workflow! ✨

FixedGetting started

Fix prompt

No emoji. Use bold only to mark a true keyword or term the reader must find—never for emphasis or decoration. Use headings only when the document has real sections. Plain text is the default.

See Signposting & Metadiscourse (structure with intent, not decoration)

13. Verb Tense Mismatch

Symptom: Unintended jumping between past, present, and future tense within a single thought or document.

Tells: "The system received the request and then processes it...", "She walked in and says..."

SmellThe user submitted the form, the API validates the input, and the database was updated.

FixedThe user submits the form, the API validates the input, and the database updates the record.

Fix prompt

Enforce verb tense consistency. Lock the document to a single baseline tense—present tense for specs, procedures, and system behavior; past tense for historical narrative or retrospectives. Fix all accidental shifts between past, present, and future tenses.

See Verb Tense Consistency

One prompt to catch them all

Prefer a single guardrail you can paste at the top of a session? This one covers the most common smells at once.

General anti-slop preamble

Write like a sharp human editor, not a chatbot. Rules for this whole session:
- No preamble, no flattery, no closing summary. Lead with the answer.
- Plain standard English. Banned: leverage, robust, seamless, delve, unlock, elevate, actionable, cutting-edge, "in the realm of", "it's worth noting".
- Commit to claims; drop filler hedges. Keep a qualifier only when the uncertainty is real, and name the exact condition.
- One term per concept—repeat it instead of finding synonyms.
- Enforce verb tense consistency; maintain a single temporal baseline (present tense for specs/behavior, past tense for retrospectives).
- Prefer strong verbs over noun phrases; name who does what.
- Prose by default; bullets only for truly parallel or sequential items.
- Ground general claims in a specific example or number.
- No emoji; bold only for true keywords.
If I am wrong, say so.