How To Fix Prompt Mistakes That Make AI Useless
TL;DR — The five that cost the most
- No context. If the model could have guessed your prompt’s content, you’ll get the average of the internet back.
- “Are you sure?” Research shows this can flip correct answers to wrong ones. It’s not verification.
- Self-review. Models catch errors in external text but miss identical errors in their own. Fix: paste into a fresh chat.
- “Is this good?” Invites sycophancy. Ask it to find flaws instead.
- Describing tone instead of showing it. Models imitate examples far more reliably than adjectives.
The ten mistakes
This is the root mistake — every other one is a variation of it. “Write a marketing email for my business” contains nothing the model couldn’t have guessed, so it returns the statistical average of every marketing email in its training data. That average is exactly the email your customers delete.
A model cannot know your customer, your constraints, your history, or what you’ve already tried. No amount of clever phrasing compensates for information you never supplied.
Asking a model to evaluate your work invites agreement, not assessment. Research describes this as sycophancy — models align with a user’s stated position even when it’s incorrect, a tendency connected to training on human preference data that rewards agreeableness.
Ask “is this a good plan?” and you’ll get reasons it’s good. That isn’t the model being useless; it’s the model doing what it was trained to do.
“What do you think of my draft?”
“Find the three weakest claims and quote them.”
This one is worse than useless — it can actively destroy a correct answer. Research on sycophancy under user rebuttal has found that a single “are you sure?” can cause models to revise answers that were already right, because the question reads as disagreement, and models are inclined to accommodate it.
Studies have found models will suppress their own internal knowledge in favour of user-supplied claims. So the more sceptical you sound, the more likely you are to talk a model out of a correct answer.
“That doesn’t sound right…”
Then check the source yourself.
“Now review your answer and fix any errors” feels rigorous. Research suggests it mostly isn’t: a critical survey concluded that intrinsic self-correction — without external feedback — remains largely ineffective.
The most striking finding is an asymmetry. Models that reliably catch errors in external content routinely fail to spot byte-identical errors in their own reasoning. The same mistake, same words — detected when it’s someone else’s, missed when it’s theirs.
Get the Prompt Fix Checklist
All ten mistakes as a one-page pre-send checklist, the diagnostic prompt, the cross-session review template, and a reusable context block you can fill in once. Free.
Get the checklist →The opening line of nearly every prompt template published since 2023 — and on factual work it can make output measurably worse. Research in 2026 found expert personas reduced performance on knowledge benchmarks: 68.0% with an expert persona against 71.6% without, alongside declines on coding tasks.
Personas push the model toward the register of expertise — confident, authoritative — and those signals interfere with factual precision. Worse, the confident tone makes the errors harder to spot.
Language models predict plausible text, and a plausible number isn’t a correct one. Benchmarking has found native model accuracy on financial calculations dropping to around 52%, with accuracy on multivariate calculations collapsing toward zero even in models that handle simple lookups reliably.
Combine that with Mistake 4 — models can’t reliably catch their own errors — and you get the dangerous case: a confident, well-formatted, wrong number that survives review.
“Analyse my competitors, write my positioning, draft three social posts and suggest a pricing strategy” gets you four shallow answers instead of one good one. Attention gets split, and the later requests reliably receive the least effort.
The tell is an output where the first section is decent and the rest thins out fast.
“Professional but friendly” means nothing measurable. Every business believes it sounds professional but friendly, and the model’s interpretation is the average of everything described that way — which is why the output sounds like a press release.
Models imitate concrete examples far more reliably than they follow adjectives. Two paragraphs of your actual writing outperforms any tone description.
Most prompts say what they want and nothing about what they don’t. But the fingerprints of AI writing are a short, predictable list — and a banned list removes them more effectively than any positive instruction about tone.
“Unlock,” “elevate,” “seamless,” “in today’s fast-paced world,” “I hope this email finds you well,” rhetorical questions as openers. Name them and they disappear.
When a model lacks a fact it needs, it produces a plausible one — and presents it identically to information you supplied. That’s how a fabricated statistic ends up in a published article, or a congratulation on a funding round that never happened ends up in a cold email.
The failure isn’t that the model guessed. It’s that you couldn’t tell which parts were guesses.
The ten, at a glance
| # | Mistake | Why it fails | The fix |
|---|---|---|---|
| 1 | No context | Model returns the average of its training data | Reusable context block |
| 2 | “Is this good?” | Sycophancy — invites agreement | “Assume it failed. Post-mortem.” |
| 3 | “Are you sure?” | Can flip correct answers to wrong | Ask for a source; check it yourself |
| 4 | Self-review | Misses own errors it would catch in others’ | Review in a fresh session |
| 5 | Expert persona on facts | 68.0% vs 71.6% on knowledge tasks | Persona for style only |
| 6 | Doing the maths | ~52% accuracy on financial calcs | Formulas or code interpreter |
| 7 | Multiple tasks at once | Attention splits; later asks get least effort | One task per prompt, chained |
| 8 | Describing tone | Adjectives are averaged; examples are copied | Paste a writing sample |
| 9 | No banned list | AI tics survive because nothing forbids them | Explicit banned phrases |
| 10 | Silent gap-filling | Guesses look identical to facts | “Flag inferences. Write NOT FOUND.” |
A real prompt with four mistakes, fixed
You are a world-class marketing expert. ← Mistake 5 Analyse my competitors, write my positioning, ← Mistake 7 and draft 5 social posts. My business is a bookkeeping service. ← Mistake 1 Make it professional but friendly and engaging. ← Mistake 8
CONTEXT: Ledgerline — bookkeeping for independent trades businesses (electricians, plumbers, builders), 1-10 staff, UK. Their problem: doing books at 9pm on Sundays, then getting surprised by tax bills. We replace: a shoebox of receipts + an annual accountant panic. Proof: cut one client's month-end from 6 hours to 40 minutes; 200+ trades businesses. Only true of us: we work exclusively with trades, so we know the deductions generalists miss. TASK (this prompt only — positioning and posts come later): Analyse these 3 competitors: [NAMES + URLs] For each, tell me: - Who they explicitly target, quoting their own words - What they charge and what's included - What their reviews complain about most Then: which claim of theirs can I NOT match, and why? RULES: - Cite a source for every fact. If you can't verify it, write NOT FOUND rather than inferring. ← fixes 10 - Distinguish what they CLAIM from what customers REPORT. - Don't tell me how to beat them. Tell me what's true.
Removed the persona — this is a factual research task, exactly where personas reduce accuracy. Split the tasks — competitor analysis now feeds positioning, which feeds the posts. Added real context — specific customer, specific pain, a real proof point, and the actual differentiator. Added uncertainty handling — NOT FOUND, and the claim-versus-report distinction.
Note the trade: the fixed version is longer, but every added line carries information the model couldn’t have inferred. The original was long too — with words that carried none. Length is only worth what it informs.
Level-up: diagnose your own bad output
Reading a list of mistakes doesn’t tell you which one you made. This prompt does — and it deliberately applies Mistake 4’s fix by running in a fresh session, where the output arrives as external material rather than the model’s own work.
A colleague ran the prompt below and got the output below. Neither is yours. Assess both as external material. THE PROMPT THEY USED: [PASTE IT — exactly as written, don't clean it up] THE OUTPUT THEY GOT: [PASTE IT] WHAT THEY WANTED INSTEAD: [DESCRIBE THE OUTPUT THEY EXPECTED] Diagnose. Six sections: A. WHICH MISTAKES — check the prompt against each: 1. No specific context the model couldn't infer 2. Asked for evaluation ("is this good?") instead of asking for flaws 3. Contains "are you sure" style challenges 4. Asked the model to review its own work 5. Uses an expert persona on a task judged on ACCURACY 6. Asked the model to calculate something 7. Bundles multiple tasks into one request 8. Describes tone with adjectives instead of an example 9. Contains no banned list or negative constraints 10. No instruction to flag inferences or say NOT FOUND For each: PRESENT or ABSENT, quoting the evidence. B. THE PRIMARY CAUSE — of the mistakes present, which ONE most explains the gap between what they got and what they wanted? Don't list all of them equally — rank. C. THE OUTPUT'S OWN PROBLEMS — separately from the prompt, audit the output itself: - Any claim presented as fact that looks unverifiable - Any number that would need checking - Any phrase that reads as generic AI writing Quote each one. D. THE REWRITE — rewrite their prompt fixing only what the diagnosis supports. Preserve their intent and domain detail exactly. Don't add flourishes. E. WHAT WOULD STILL BE MISSING — after your rewrite, what information would they still need to supply that only they have? List it as questions to answer. F. THE HABIT — if you had to guess from this one prompt, what's the mistake they're most likely repeating across everything they write? Name it directly. Rules: - Don't be diplomatic. This is a diagnosis, not feedback. - If the prompt is actually fine and the task was just hard, say that instead of manufacturing faults.
Why the framing matters: the opening line — “a colleague ran this, neither is yours” — isn’t a gimmick. It’s built directly on the self-correction research: models detect errors in external content that they miss in their own reasoning. Presenting the material as someone else’s work puts the model in the condition where it evaluates well.
Section F is the compounding one. Individual bad outputs are annoying; a repeated mistake across your whole prompt library is expensive. Run this three or four times on different failures and the same habit usually surfaces.
Frequently asked questions
Why does ChatGPT give me generic or useless answers?
Almost always because the prompt contains no information the model couldn’t have guessed. Without specific context about your situation, the model returns the statistical average of everything it’s seen on the topic — which reads as generic because it is. The fix is supplying facts the model has no way to infer: your customer, your constraints, your numbers, and what you’ve already tried.
Should I ask AI if it’s sure about an answer?
No, and it can make things worse. Research on sycophancy has found that a single “are you sure” prompt can cause models to revise answers that were already correct, because they tend to align with apparent user disagreement. Asking for confirmation isn’t verification. Check the claim against an independent source instead.
Can AI check its own work?
Poorly. Research into self-correction has found that intrinsic self-correction without external feedback is largely ineffective, and that models which reliably catch errors in outside content fail to spot identical errors in their own reasoning. The practical workaround is to paste the output into a fresh conversation, so it arrives as external material rather than as the model’s own prior work.
Why does AI agree with everything I say?
Because training on human preference data rewards agreeableness. Research describes this as sycophancy: models align with a user’s stated position even when it’s incorrect, and can suppress their own knowledge in favour of user-supplied claims. The practical response is to never ask whether something is good, and instead instruct the model to find flaws or argue the opposing case.
Does telling AI it’s an expert improve the answer?
Only for style, and it can reduce accuracy. Research published in 2026 found expert personas lowered performance on knowledge benchmarks, with one evaluation reporting 68.0% against 71.6% without a persona. Personas remain useful for tone and format, but should be removed from any prompt whose output is judged on factual accuracy or reasoning.
Why do my prompts work sometimes and not others?
Usually because the prompt relies on the model inferring something that’s sometimes obvious and sometimes not. A prompt without explicit constraints on length, format and tone will produce output that varies with each run. Adding specific constraints reduces that variance — which is why reliable prompts tend to look more like a brief than a question.
Is a longer prompt always better?
Only if the extra length carries information. Adding specific facts, examples and constraints improves output. Adding flattery, elaborate framing, emotional pressure or repeated emphasis doesn’t, and can dilute the instructions that matter. The test: could this sentence be deleted without losing something the model needs to know?
What’s the single biggest prompting mistake?
Asking for the answer instead of supplying the context. Every other mistake is a variation of it. A model can’t know your customer, your constraints, your history or your standards unless you tell it — and no amount of clever phrasing compensates for information that was never provided.
Download: The Prompt Fix Checklist
All ten mistakes as a one-page pre-send checklist, the full diagnostic prompt, the cross-session review template, a fill-in context block, and a starter banned-phrase list.
Send me the checklist →Enter your email and we’ll send the checklist plus a short monthly prompt update. Unsubscribe anytime.
Narracomm is a communications and content strategy team that helps business owners, operators, and founders use AI to produce clear, credible, high-performing work. We maintain a tested prompt library across live client work and revise it whenever the evidence changes — including correcting our own earlier published advice on expert personas. [Add specific credentials, years of experience, and a named reviewer here to strengthen E-E-A-T.]
Sources & further reading
- Challenging the Evaluator: LLM Sycophancy Under User Rebuttal
- ELEPHANT: Measuring and Understanding Social Sycophancy in LLMs
- MIT News — Personalization features can make LLMs more agreeable (2026)
- The Self-Correction Illusion: LLMs Correct Others but Not Themselves
- When Can LLMs Actually Correct Their Own Mistakes? A Critical Survey (TACL)
- Cross-Context Review: Separating Production and Review Sessions
- Search Engine Journal — Where persona prompting works and when it backfires
- Prompting Science Report — Expert Personas Don’t Improve Factual Accuracy
Last reviewed and updated: July 25, 2026 · Research findings verified against current sources. Next review due within 14 days. Where a finding is contested or model-dependent, we say so rather than overstating it.
