Prompt engineering guides are usually 5,000 words of theory. This one is different: 10 copy-paste templates that fix the most common AI failures — vague answers, generic writing, hallucinated facts, and ignored instructions. Each template includes why it works and where to customize. They work in ChatGPT, Claude, and Gemini alike.
Why Templates Beat Theory
Every reliable prompt does some combination of four jobs: give context, constrain the output, show an example, and define success. The templates below are pre-built combinations of those four moves. Steal them, swap the bracketed parts, and you'll outperform 90% of users without learning any theory.
The 10 Templates
1. The Role + Audience Lock (fixes generic answers)
You are a [senior tax accountant]. Explain [topic] to [a freelancer with no finance background]. Use concrete numbers in every example. Avoid jargon — where a technical term is unavoidable, define it in one sentence.
Why it works: pinning BOTH the expert role and the audience kills the wishy-washy middle-ground tone models default to.
2. The Format Contract (fixes rambling)
Answer in exactly this structure: 1) One-sentence answer. 2) Three bullet points of reasoning. 3) One risk or exception I should know. Nothing else.
Why it works: models follow explicit output contracts remarkably well — and "nothing else" suppresses the padding.
3. The Anti-Hallucination Clause (fixes made-up facts)
Answer using only information you are confident about. For anything uncertain, write "UNVERIFIED:" before the claim. If you don't know, say "I don't know" — do not guess.
Why it works: giving the model an explicit permission structure for uncertainty dramatically reduces confident fabrication. Pair with a follow-up: "List the three claims above most likely to be wrong."
4. The Example Anchor (fixes wrong style)
Rewrite the following in the same style as this example. Example: "[paste 2–3 sentences of the style you want]". Now rewrite: "[your text]". Match sentence length, tone, and vocabulary level of the example.
Why it works: one concrete example outperforms ten adjectives. "Professional but friendly" means nothing; a sample paragraph means everything.
5. The Iteration Loop (fixes first-draft mediocrity)
Draft [the thing]. Then critique your own draft against these criteria: [clarity / persuasiveness / accuracy]. Then produce an improved version incorporating your critique. Show only the final version.
Why it works: self-critique passes measurably improve output, and hiding the intermediate steps keeps the response clean.
6. The Decision Forcer (fixes fence-sitting)
Compare [option A] and [option B] for [my specific situation]. End with a one-line recommendation: which one, and the single strongest reason. You must pick one — "it depends" is not an acceptable answer.
Why it works: models hedge by default. Explicitly banning the hedge produces the committed analysis you actually wanted.
7. The Extraction Grid (fixes messy data pulls)
From the text below, extract every [company name, price, date]. Output a markdown table with columns: [X / Y / Z]. If a value is missing, write "—". Text: [paste].
Why it works: naming the columns and the missing-value rule eliminates the two ways extraction usually breaks.
8. The Beginner's Ladder (fixes overwhelming explanations)
Explain [complex topic] three times: first to a 10-year-old in 2 sentences, then to a college student in 5 sentences, then to a practitioner in one paragraph. Label each level.
Why it works: the laddered format forces the model to isolate the core idea before adding complexity — and you choose the level you need.
9. The Devil's Advocate (fixes confirmation bias)
Here is my plan: [plan]. Argue against it as a skeptical expert would. Give the three strongest objections, each with the specific scenario where my plan fails. Then rate how fatal each objection is (1–10).
Why it works: models agree with users by default. Assigning the opposition role is the only reliable way to get real criticism — use it before any big decision.
10. The Reusable System Prompt (fixes repeating yourself)
For this entire conversation: I am a [role] working on [project]. Always [format preference]. Never [pet peeve]. My audience is [audience]. Confirm you've understood in one line, then wait for my first task.
Why it works: front-loading standing instructions once beats re-explaining in every message — and works in ChatGPT custom instructions, Claude projects, and Gemini gems alike.
Combining Templates
The real power move is stacking: #1 + #2 + #3 is a reliable default for factual work ("expert role + strict format + no guessing"). #4 + #5 is the writing combo. #6 + #9 is the decision-making pair.
And if you're automating prompts across models via API, templates make even more sense — the same structures work whether the backend is GPT, Claude, or a cheaper routed model, letting you downgrade to budget models without quality collapse. For agent-style tasks that browse and act, see our ChatGPT Agent Mode recipes — the task-brief formula there is template #2's big brother.
Save this page, steal the templates, and stop writing prompts from scratch.