Should I give ChatGPT an example of what I want?
Give it a real example or none at all, and attach the example to the input rather than to the answer. Across all 455 prompts we publish, not one contains a sample of the finished answer inside the prompt text. What they carry instead is 988 exampled input slots out of 3,263, and 33 prompts that make you paste a document you already have.
That is not what most advice says, and it is not what the library looks like from the outside either. Every prompt we sell ships with a worked example underneath it. There are 455 of them and they run a median of 51 words. But every single one sits outside the prompt, in the pack, for you to read. Zero of them are inside the text you paste into ChatGPT.
Once you notice that split, the rule behind it is easy to state. An example is a sample of a population, and the model infers the population from the sample. Give it a real sample and it infers your actual work. Give it one you invented to demonstrate what you meant, and it infers your invention, then hands you a tidier copy of it.
What do 455 working prompts actually do with examples?
They example the inputs heavily and the output not at all. Here is the whole library in one table.
| What was measured | Across all 455 prompts |
|---|---|
| Prompts containing a sample of the finished answer | 0 of 455. No prompt body contains an Example Output block or anything playing that role. |
| Bracketed input slots | 3,263 in total. 988 carry an inline example of what to type there. |
| Prompts with at least one exampled input slot | 360 of 455. |
| Prompts that make you paste a real document | 33 of 455, through 35 paste slots, spread across 13 of the 15 packs. |
| Prompts quoting a phrase the model must not use | 64 of 455, quoting 158 banned phrases in total. |
| Longest quoted specimen anywhere inside a prompt body | 73 words, and it is a list of things you supply, not an answer. |
| Worked examples shipped with the prompts, outside the prompt text | 455, one per prompt. Median 51 words, range 21 to 158. |
Read the first and last rows together. Every prompt in the library comes with a finished example, and not one of those examples is ever shown to the model. They exist so a buyer can tell, before typing anything, what the prompt is going to produce. They are documentation, not instruction.
Where does the example actually sit?
Inside the input slot it explains. This is the standard form, taken verbatim from our Sensory Product Description prompt in the Ecommerce pack:
Write a product description for [PRODUCT NAME], a [PRODUCT TYPE, e.g., "handmade soy candle" or "ergonomic office chair"]. Target customer: [IDEAL BUYER, e.g., "remote workers who spend 8+ hours at a desk"]. Price point: [PRICE, e.g., "$89"]. Key differentiator: [WHAT MAKES IT DIFFERENT, e.g., "adjustable lumbar support with memory foam"].
Four examples in six lines, and every one of them is answering a different question from the one people usually think they are answering. They are not saying here is the kind of description I want. They are saying here is the resolution at which you should describe your own buyer. The example calibrates your typing, not the model's writing.
Exactly one prompt in the library sets an example off in its own labelled block, and it is worth quoting in full because the label is the thing people expect to mean output. It is our Panel Interview Prep prompt:
I have a panel interview with these interviewers: [LIST EACH INTERVIEWER: NAME, TITLE, AND WHAT YOU THINK THEY CARE ABOUT] Example: - Sarah Chen, VP Engineering — cares about technical depth and system design - Mike Rodriguez, Product Lead — cares about collaboration and user focus - Anya Patel, HR Business Partner — cares about culture fit and growth mindset Role: [JOB TITLE] For each interviewer, generate: 1. The 2 questions they're most likely to ask based on their function 2. What they're really evaluating (the hidden rubric)
The one block in 455 prompts that is literally headed Example is an example of the list you paste in. It sits directly under the slot it belongs to and directly above the deliverables, which are described rather than demonstrated.
Why does an invented example make the answer worse?
Because you are handing over a specimen and a claim at the same time, and the claim is the part you did not mean.
We are labelling this section as reasoning rather than measurement, because that is what it is. When you write something like: Boost Your Morning Routine in 5 Minutes, you intend it as a gesture at a category. It arrives as a data point. The model has one instance of the target class and no way to know which of its properties are load bearing. Was the imperative verb the point, or the number, or the time frame, or the colon, or the six word length? All of them are equally present in the sample, so all of them get preserved, and you receive five headlines that differ from yours mainly in vocabulary.
The second cost is that the example silently overrides your description. If you write a careful paragraph about wanting a warm and unhurried tone, and then paste a snappy invented sample, the sample wins. Specifications are abstract and examples are concrete, and concrete beats abstract when the two disagree. Most people who feel ignored by an AI wrote both and only noticed one.
Real examples do not have this problem, because you did not manufacture them to make a point. Three of your actual published posts carry the properties your writing genuinely has, in their real proportions, including the ones you would never have thought to name.
When is an example worth including?
The test is whether the example already existed before you needed it.
| What you have | Send it? | Why |
|---|---|---|
| Your own published work, in the voice you want | Yes, paste it | It carries properties you cannot name. This is what our voice and style prompts are built around. |
| The document being worked on: a resume, a contract, a product page | Yes, paste it | 33 of our 455 prompts do exactly this, through 35 paste slots in 13 of the 15 packs. |
| A competitor's page or a template you are matching | Yes, paste it | Real artifact, real constraints. Say explicitly what you are matching and what you are not. |
| A description of the kind of input you will supply | Yes, inline | 988 of our 3,263 input slots carry one. It sets the resolution you are expected to answer at. |
| A phrase you never want to see | Yes, quote it | 64 of our prompts quote 158 banned phrases. A negative example removes one option and leaves the rest open. |
| A sample of the answer, written by you, just now, to illustrate | No | It becomes the target. Describe the answer's shape instead, the way 276 of our prompts do across 1,469 numbered lines. |
| A half remembered example of something you saw once | No | You will reconstruct it wrong and then be matched against your reconstruction. |
Five of those seven rows are yes. This page is not an argument against examples. It is an argument against the one kind of example that is almost always invented, which happens to be the kind people mean when they ask the question.
What do these prompts do instead of showing an answer?
They specify it. Three mechanisms, all countable, all copyable.
They enumerate the deliverables. 276 of the 455 prompts list what comes back as a numbered set, 1,469 lines in total, at a median of 5 items among the prompts that use a list. Each line names a thing rather than describing a quality, which is why the list can be checked against the answer afterwards.
They set hard limits. 150 of the 455 carry an explicit number: a word count, a character count, a sentence count or a fixed number of bullets. Length: 150-200 words does the job a sample paragraph was going to do, without smuggling in a topic, a tone and a sentence rhythm alongside the length.
They name what is forbidden. 64 of the 455 quote phrases the model may not use, 158 in total. This is the one place the library does show the model specimen text, and the asymmetry is the interesting part. Here is a banned list verbatim from our Investor Pitch prompt in the Startup Founder pack:
Banned phrases: "We're disrupting," "Our platform leverages," "In today's world," "We're passionate about," "Think of us as Uber for X."
Five specimens, every one of them a negative. Quoting a phrase you do not want subtracts a single option from a very large space. Quoting a phrase you do want multiplies everything that resembles it. The library leans hard on the first and never uses the second, and once you have seen that stated plainly it is difficult to go back to pasting invented samples.
Then what are the Example Output blocks under every prompt for?
You, before you buy. Not the model, after you paste.
All 455 prompts ship with one and they average out at a median of 51 words, from a 21 word minimum to a 158 word maximum. They answer a buyer's question, which is whether this prompt produces something worth the money, and they answer it in the space of a glance. That is a completely different job from constraining a generation, and they are physically outside the fenced prompt body precisely so nobody confuses the two.
Worth saying plainly, because it is the obvious objection: the fact that our prompts are built this way is evidence about how we write prompts, not proof about how language models behave. We say the same thing on every page in this series. The count is real, the sample is 455 prompts written for paying users across 15 professions by people whose incentive is that the prompt works first time, and it is still a census rather than a controlled test.
A prompt you can copy
This one is for the common case where you do have real material and you want a reusable specification out of it, rather than pasting the same three documents into every future prompt. Feed it your actual work and it gives you back the spec that work implies.
I want to turn some real examples of my own work into a written specification I can reuse in future prompts, so that I stop pasting the examples themselves every time. Here are [NUMBER, e.g., "3"] real pieces I want future output to resemble. They are genuine finished work, not written for this exercise: --- PIECE 1 --- [PASTE] --- PIECE 2 --- [PASTE] --- PIECE 3 --- [PASTE] What I want future output for: [THE JOB, e.g., "weekly customer update emails for a 40 person SaaS company"]. What in these pieces is deliberate: [ANYTHING YOU KNOW YOU DO ON PURPOSE, e.g., "I always open with the problem, never with a greeting", or "unsure"]. Do this, and do not write anything in my style yet: 1. Describe what these pieces have in common, as a numbered list of rules I could hand to someone who has never read them. Cover structure, opening move, sentence length, vocabulary level, what I do with numbers, how I close, and anything else that recurs in all three. 2. For each rule, quote the shortest fragment from the pieces that shows it, so I can check the rule against the evidence. 3. Separate the rules that hold in all three pieces from the ones that hold in only two. Label them Confirmed and Provisional. Do not merge them. 4. List what you could NOT determine from three samples, and say what kind of extra material would settle each one. 5. Write the specification as a block I can paste into future prompts. Use explicit numbers wherever the pieces support one, such as typical word count and typical sentences per paragraph. 6. Write a short banned list: phrases or moves that appear nowhere in my three pieces but that an AI would reach for by default on this job. Do not invent a rule to make the list look complete. If these three pieces only support four rules, give me four.
The output is a specification, and a specification is the thing you actually wanted when you were about to paste an invented example. Step 6 is the one people leave out, and it is the step that produces the banned list that 64 of our own prompts carry.
How these numbers were measured
The figures are direct counts over the 455 prompts published in our 15 packs, taken on 17 September 2026 by parsing every fenced prompt body out of the pack files. Prompts per pack run from 28 to 35, and all 455 parsed cleanly.
The zero is the number most worth checking, so here is how it survived three attempts to break it. A narrow first pass for output sample language found 3 candidates, and all 3 were read in full: two were looks like this: followed by a paste slot for the reader's own training program and daily schedule, and the third was the Panel Interview block quoted above. All three are inputs. A second, deliberately wide pass matched any prompt line containing example, sample, template, modeled on, mimic or like this, which caught 66 prompts and 79 lines, and every line was read. They fall into three groups: deliverables that happen to be templates the model is asked to produce, instructions to include a concrete example in the answer, and input slots. A third pass measured every double quoted fragment inside every prompt body: 2,354 fragments, the longest 73 words, and that longest one is a list of nine image concepts the user supplies. Against a median finished deliverable of 51 words, no prompt body contains a specimen of the answer at anything close to full size. Separately, the string Example Output appears in 0 of 455 prompt bodies.
Input slots were counted as bracketed placeholders, and a slot counts as exampled when it contains e.g. inside the brackets, which is a strict test that ignores slots offering a list of options. Paste prompts were matched on a literal paste placeholder and the resulting list of 33 was read by name. Banned phrases were counted as quoted strings on a line carrying an explicit prohibition, and 12 of the 64 were spot checked by reading the prompt.
The limit is the usual one and it is real. This is a census of prompts we believe work, not a controlled test of a prompt with an example against the same prompt without one. We did not run that test and nothing above should be read as though we did. The mechanism in the third section, about an invented specimen overriding a written description, is reasoning and is labelled as such where it appears. The one place we have run a controlled comparison on a prompt phrasing, with the markers fixed before either output existed, is the persona test in does telling ChatGPT to act as an expert actually work.
Questions people ask about giving ChatGPT examples
Give it a real example or none at all, and attach the example to the input rather than to the answer. Across the 455 prompts we publish, not one contains a sample of the finished answer inside the prompt text. What they do carry is 988 exampled input slots out of 3,263, and 33 prompts that make you paste a document you already have. If the thing you want does not exist yet, do not invent a specimen of it. Describe its shape instead, which is what 1,469 numbered requirement lines across 276 of our prompts do.
One real one per input, and zero invented ones of the output. In our library the median prompt carries 2 inline examples and they are spread across its input slots rather than stacked on the answer. The reason is that a second invented example does not add information, it adds agreement: two specimens you made up look like a pattern, and the model will extract that pattern and hand it back. Two real specimens genuinely do help, and that is the technique usually called few shot prompting, but it needs real labelled pairs from your own material, not two guesses at what good might look like.
Then specify the output instead of illustrating it. Our prompts do this in three ways you can copy directly. 276 of 455 list the deliverables as numbered items, 1,469 lines in total. 150 of 455 set an explicit limit such as a word count, a character count or a number of bullets. And 64 of 455 quote the exact phrases the model must not use, 158 banned phrases in total. A specification constrains the answer without narrowing it onto one invented specimen, which is exactly what an imaginary example does.
Yes, and that is the point when the example is real and the problem when it is not. An example is a sample of a population, and the model infers the population from the sample. If the sample is three of your actual published posts, the inferred population is your writing and the resemblance is the benefit. If the sample is a paragraph you wrote thirty seconds ago to demonstrate what you meant, the inferred population is that paragraph, and you get a paraphrase of your own rough draft back with better grammar.
Yes, and our library does this far more readily than it shows good examples. 64 of the 455 prompts quote at least one phrase the model is forbidden to use, 158 phrases in total, in lines such as Banned phrases: We are disrupting, Our platform leverages, In today's world. A negative example is safe in a way a positive one is not. Naming a phrase you do not want removes one option and leaves every other option open, while showing a specimen you do want closes the field down to things that resemble it.
Paste it, whenever it exists. 33 of our 455 prompts hand the model a real artifact through a paste slot, 35 slots in total, spread across 13 of the 15 packs. They take your actual resume, your current training program, your existing product page copy, the contract text, the job description. None of them says describe your resume, because a description of a document is your summary of it, and the model then works from your summary rather than from the thing. The one cost is length, and our prompts absorb it: the median prompt is 174 words before you paste anything into it.
Next to the thing it is an example of. In our library the examples sit inside the bracketed input slot they clarify, in the form PRODUCT TYPE, e.g., handmade soy candle or ergonomic office chair. That placement is deliberate: the example is answering the question what do you want me to type here, so it belongs where you are typing. Examples parked in a block at the top of a prompt get read as a template for the answer whether you meant them that way or not, which is the failure this whole page is about.
Related reading and next steps: the biggest legitimate use of a real example is matching your own writing, and how to get ChatGPT to write in your voice covers turning your material into a reusable style sheet. For the facts that go in alongside the examples, what information should I give ChatGPT counts every input slot in the library. If you were reaching for an example because the shape kept coming back wrong, how to get ChatGPT to follow the format you asked for is the direct fix, and how to tell ChatGPT what not to do covers the banned list in depth. If the output reads like the average of a category, that is measured in why ChatGPT gives generic answers, and for how much of all this belongs in a single message, see should I ask ChatGPT one thing at a time and how long should a ChatGPT prompt be. To start from prompts that already do this, browse the prompt packs or read the how to use guide.