What information should I give ChatGPT to get a good answer?

Published 9 September 2026 by PrecisionPrompts. Every figure on this page is a direct count over the 455 prompts we publish across 15 packs, measured on that date.

Give it the facts it would otherwise invent: who the answer is for, a number that bounds it, the specific named subject, and the tone or channel it has to fit. Across the 455 prompts we publish, the median prompt asks the user to supply 7 specific pieces of information, and 987 of those 3,262 requests carry a filled example rather than just a label.

The complaint that brings people here is usually phrased as a quality problem. The answer is fine but bland, or it is on the right topic but written for nobody in particular, or it is confidently wrong about a detail nobody supplied. None of those is the model being weak. They are the model resolving a question that was left open, and it resolves an open question the same way every time: with the most average answer available.

So the practical question is not "how do I write a better prompt", it is "which facts am I leaving the model to guess". What follows is a census of our own library on exactly that: what 455 prompts that people bought and kept using ask their user to hand over, and in what form. It describes a house style that survived contact with buyers. It is not a controlled test of whether supplying a given fact improves an answer, and where that limit matters we say so.

What do working prompts actually ask you for?

Seven things, over and over, and the first four appear in every pack we publish. That last part is the finding worth keeping: the same categories recur whether the task is a property listing, a lesson plan or a licence agreement, which makes this a checklist rather than a genre convention.

Read the top row and the bottom row together and the pattern is clear enough to act on. The inputs that dominate are the ones that close a decision the model would otherwise make silently. A number closes the size of the answer. An audience closes its vocabulary and its objections. A named subject closes what the answer is even about. The instruction people reach for first, a list of things not to do, sits at the bottom on 34 percent, because a prohibition only helps once the model already knows what it is making and for whom.

The audience row deserves a line of its own. It is the single most under-supplied input in the prompts people write for themselves, and it is doing more work than it looks: audience sets reading level, vocabulary, which objections get answered and roughly how long the answer runs, all at once and all without being asked. Leave it out and the model writes for a general business reader. That register is precisely what people are describing when they say an answer sounds like AI, which we covered separately in why ChatGPT gives generic answers.

How much information is enough?

About seven facts. Not three, and not a page of background.

Seven is a median, not a target, and the spread around it is the more useful signal. The count tracks how many decisions the task actually contains. Our AI Image Prompts pack averages 4.3 input slots per prompt, because a picture has few open decisions once you have named the subject, the surface and the light. The Content Creator pack averages 9.7 and the Legal Professional pack 9.5, because a piece of content has an audience, a platform, a length, a hook, a call to action and a voice, and a contract has parties, terms, caps, cure periods and a governing state.

So the honest rule is: count the decisions in the task, then supply that many facts. If you cannot name a fifth decision, you do not need a fifth fact, and padding the prompt with background nobody asked for is the failure mode measured in how long should a ChatGPT prompt be. The median prompt in our library runs 1,129 characters, roughly a screen. Seven facts fit in a screen comfortably when each is a labelled line rather than a paragraph.

Why naming the input is not enough

This is the part people skip, and it is the part that changes the answer most. Naming an input tells the model which fact you are supplying. It does not say how specific that fact is meant to be, and specificity travels in both directions.

360 of the 455 prompts, 79 percent, carry at least one filled example. They are concentrated exactly where you would want them: the bare labels sit on slots where only one kind of answer is possible, like a client name or a state, and the filled examples cluster on the slots where the model needs to know the grain.

The mechanism is worth stating plainly, because it is the whole reason this works. A slot that reads "target customer" and shows nothing gets filled in by the user at whatever grain feels natural, which is usually a category: "small businesses". A slot that reads target customer and shows "remote workers who spend 8+ hours at a desk" makes the user supply something at that grain, and the answer inherits it. The example is not mainly there to teach the model. It is there to raise the specificity of what you type in.

What a fully specified input block looks like

Here is one from our Legal Professional pack, reproduced complete. It runs 1,147 characters, close to the library median, and it carries 12 input slots of which 9 include a filled example.

Draft a Master Services Agreement where [PROVIDER NAME] ("Provider") will
deliver [TYPE OF SERVICES, e.g., "custom software development and ongoing
maintenance"] to [CLIENT NAME] ("Client").

Key commercial terms:
- Payment: [PAYMENT STRUCTURE, e.g., "monthly retainer of $X with net-30 terms"]
- SOW process: new work scoped via Statements of Work attached as exhibits
- IP ownership: [WHO OWNS DELIVERABLES, e.g., "Client owns all custom
  deliverables; Provider retains pre-existing IP with perpetual license to
  Client"]
- Liability cap: [CAP AMOUNT OR FORMULA, e.g., "12 months of fees paid"]
- Indemnification: mutual, with specific carve-outs for [SCENARIOS, e.g., "IP
  infringement, gross negligence, data breach"]
- Termination: [NOTICE PERIOD, e.g., "30 days written notice"] for convenience;
  immediate for material breach with [CURE PERIOD, e.g., "15-day cure period"]
- Data handling: Provider will process [TYPE OF DATA, e.g., "customer PII"]
  subject to [APPLICABLE REGULATIONS, e.g., "CCPA and SOC 2 requirements"]
- Governing law: [STATE]

Include a severability clause and an order-of-precedence clause (SOW > MSA in
case of conflict).

Three things are happening in that block, and all three are copyable into any task.

1. The facts are labelled lines, not a paragraph. "Payment", "Liability cap", "Governing law". A labelled line is a slot the model can see is filled or empty. A paragraph of context is a slot it has to parse first.
2. The examples set the grain. "12 months of fees paid" tells the user that a liability cap is expected as a formula, not as a vague preference. Nine of the twelve slots do this.
3. The parts with no real decision are written as facts, not slots. The SOW process and the severability clause are stated outright, because there is nothing for the user to choose. Turning those into slots would add work without closing a decision.

The same shape holds in a completely different domain. This one is from the AI Image Prompts pack, 8 slots and 7 filled examples, doing the identical job for a picture.

Emblem-style badge logo for "[BRAND NAME]", a [BUSINESS TYPE, e.g.,
"small-batch craft coffee roaster"]. Circular or shield shape containing: the
brand name set along [TEXT PLACEMENT, e.g., "the inner top curve of the
circle"], a central illustration element of [ILLUSTRATION, e.g., "a single
coffee cherry branch with three berries, line-drawn"], and [SECONDARY TEXT,
e.g., "Est. 2024" along the bottom curve]. Style: [AESTHETIC, e.g.,
"hand-drawn artisan with slight line weight variation, suggesting letterpress
or woodcut"]. Two-color maximum: [COLOR 1, e.g., "espresso brown #3C2415"] and
[COLOR 2, e.g., "cream #F5F0E8"]. Vintage craft aesthetic, not corporate.

What if I do not have all of it?

Say the fact is missing rather than deleting the line. Our library does this 22 times, phrased as "provide what you have", "or none", "if known", "[IF APPLICABLE]" and "leave blank".

The reasoning is small and slightly counterintuitive. Deleting an input line removes the decision from the prompt, which hands it straight back to the model, and the model fills it silently and never mentions it. Writing "conversion rate: unknown" keeps the decision visible and changes the behaviour: the model will work around the gap, flag it, or ask. That is the difference between an answer with a stated hole in it and an answer with an invented number in it, and only one of those is safe to paste into a document.

The related case is material you already have. 39 of our prompts ask the user to paste in something real: a current draft, an existing listing, a job description, a transcript. If you have the artifact, paste the artifact. A description of your existing copy is a summary of it, and the model will work from your summary rather than from the thing itself.

Does giving more information always help?

No, and the library is fairly clear about where the ceiling sits. 88 prompts ask for ten or more inputs, but the maximum anywhere across 455 prompts is 24, and the prompts at that end are the ones with genuinely many independent variables, like a multi-party agreement or a full onboarding audit. Nothing in the library asks for forty facts, because past a point the extra facts are not closing decisions. They are context the model has to hold while answering.

The test is mechanical. For each fact you are about to supply, name the decision it closes. "Our funding stage is seed" closes nothing if the task is writing a product description. "Price: $89" closes the entire question of which price framing arguments are even available. If you cannot name the decision, the fact is not helping. The same logic applies to prohibitions, which we go into in how to tell ChatGPT what not to do.

A prompt that tells you which facts you left out

Paste in a prompt that keeps producing bland or wrong answers. This works out which decisions you left open, which is usually more useful than a rewrite.

Here is a prompt I keep using, and the answers come back too generic:
[PASTE YOUR PROMPT]

Do not rewrite it yet. First, audit it. Produce:

1. **Decisions left open**: every choice you would have to make yourself in
   order to answer this prompt, as a numbered list. For each one, say what you
   would default to if I did not tell you.
2. **Missing facts, ranked**: which 3 of those decisions would change the
   answer most if I closed them. Rank by how much the output would differ.
3. **A fill-in block**: rewrite my inputs as labelled lines, one fact per
   line, and for each line include a filled example of a good value in the
   form [LABEL, e.g., "a real value at the right level of detail"]
4. **What I should not add**: anything I already supplied that closed no
   decision

Be specific about the defaults. "I would assume a general business audience"
is the useful answer. "It depends" is not.

Item 1 is the one to read carefully. The list of defaults it produces is a direct readout of every place your prompt was ambiguous, and it is usually shorter and more embarrassing than expected.

How these numbers were measured, and what they do not show

Every figure here is a direct count over the 455 prompts published in our 15 packs, taken on 9 September 2026. An input slot is a bracketed placeholder in the prompt body, and there are 3,262 of them across the library. Prompt bodies were read from the published pack files, and code fences shorter than 40 characters were excluded so that inline snippets are not counted as prompts.

These are counts over prompts we publish and believe work. They describe a house style that survived contact with buyers, and they are not a controlled comparison of a prompt with an input against the same prompt without it. Nothing here should be read as "supplying an audience measurably improved the answer", because we did not run that test. Where we have run a controlled test, on a single build with the markers fixed in advance, it is reported separately in does telling ChatGPT to act as an expert actually work, including the part of the result that went against us.

Table 1 is a keyword match and has a known blind spot in both directions. A prompt that asks for an audience in unusual wording is not counted, and a prompt that mentions a date in passing may be. Those category totals are reliable to a few percent, not to the unit. Tables 2 and 3 are structural counts over bracketed slots and are exact, subject to the one judgement call in Table 3 about what separates a description of a value from a filled example.

Questions people ask about what to tell ChatGPT