Should I ask ChatGPT one thing at a time, or put it all in one prompt?
Put everything about one job in one prompt, and start a new prompt when the job changes, not when the list gets long. Across all 455 prompts we publish, 454 are written to be answered in a single turn. Exactly one asks the model to stop and wait for you, and that one is a mock interview, where the back and forth is the product rather than the delivery method.
This surprises people, because the instinct runs the other way. Asking for four things at once feels greedy, and asking for them one at a time feels careful. The library says the opposite. The median prompt in it hands over 7 facts and asks for 4 separate things back, in one message, and then stops talking.
What follows is the census, why the careful looking approach quietly costs you, the three things that actually do earn a second prompt, and what that second prompt has to contain.
What do 455 working prompts actually do?
They front load everything and expect one answer. Here is the shape of the whole library in one table.
| What was measured | Across all 455 prompts |
|---|---|
| Turns the prompt expects | One, in 454 of 455. A single prompt tells the model to pause and wait for a reply. |
| Length | Median 174 words. Range 73 to 424. |
| Facts supplied up front | 3,239 bracketed input slots in total. Median 7 per prompt, from 1 to 24. |
| Things requested back | 1,469 numbered deliverable lines in total. 276 prompts use a numbered list, at a median of 5 items and a range of 2 to 11. |
| Prompts asking for two or more different kinds of output at once | 91 of 455, for example a diagnosis and a rewrite and a changelog in one answer. |
| Prompts that continue earlier work | 32 of 455, and all 32 paste that earlier work back in rather than referring to it. |
Read the first and last rows together, because they are the finding. The library almost never has a conversation, and on the rare occasion it continues one, it carries the previous material with it instead of trusting the chat to remember.
Here is what packing a job into one turn looks like. This is our Product Page Tear Down prompt from the Ecommerce pack, shortened only by trimming the rules block at the end:
I'll paste my current product page copy below. Tear it apart and rewrite it to convert better. Current copy: --- [PASTE YOUR ENTIRE CURRENT PRODUCT PAGE COPY HERE] --- Product: [PRODUCT NAME]. Price: [PRICE]. Target customer: [PERSONA]. Current conversion rate (if known): [RATE]. Top traffic source: [SOURCE]. Provide: 1. **Diagnosis**: bullet-point list of what's wrong with the current copy (be blunt). For each issue, cite WHY it hurts conversion. 2. **Rewritten copy**: full replacement, section by section. 3. **Changelog**: a table showing "Original | Rewrite | Why changed" for every major change. 4. **3 quick wins**: the smallest changes that would have the biggest impact.
That is four genuinely different artifacts: a critique, a replacement, a comparison table and a shortlist. The obvious way to get them is four messages. The prompt asks for all four at once, and the reason is visible in item 3. The changelog is a comparison between items 1 and 2, so it can only be written by something holding both at the same time. Split this into four messages and the changelog gets reconstructed from memory instead of produced alongside the thing it documents.
Why does breaking it up feel safer?
Because each small message is easier to write, and the cost of splitting does not show up until later, in a form that does not look like it came from splitting.
Here is the mechanism, and we are labelling it as reasoning rather than measurement, because that is what it is. Every message after the first is answered in the shadow of the one before it. Ask for the outline, then the intro, then the body, then the conclusion, and the intro is written without knowing how the piece ends, while the conclusion is written to fit an intro the model has already committed to. A requirement you mention in message four cannot reach backwards and change message two. So you either live with the seam or you ask for a rewrite, and the rewrite is a fifth message that now has four earlier answers pulling on it.
A single prompt naming all four parts lets the model plan across them before it commits to any one of them. That is the same reason the tear down prompt above can ask for a changelog at all.
The second cost is subtler. In a long chain, the earliest answer is the one with the least context behind it, and it is also the one everything after it has been built to match. A wrong turn taken in message one does not get corrected by message six. It gets defended by it.
So when is a second prompt the right call?
Three situations. None of them is that your request got long.
| Situation | One prompt or two? | Why |
|---|---|---|
| You want four parts of one deliverable | One | The parts have to agree with each other. 276 of our prompts enumerate their parts in a single numbered list, at a median of 5. |
| You want several different artifacts about the same material | One | 91 of our prompts do exactly this. The material is supplied once and the artifacts stay consistent with each other. |
| The request feels long or demanding | One | Length is not the constraint. Our longest published prompt is 424 words and it is a single turn. |
| The subject changes | Two | The customer emails and the quarterly budget are two jobs. Nothing in the second is improved by the first being in the window. |
| You cannot write the next brief until you have read this answer | Two | A real dependency, not a preference. Our 32 teardown and rewrite prompts assume exactly this. |
| The input is bulky enough to bury the instruction | Two | A 3,000 word document plus a second unrelated task makes the actual ask hard to locate. |
The line those six rows draw is between splitting by step and splitting by subject. Splitting by step is what people do by default, and it is the one that costs you. Splitting by subject is nearly always right, and it is cheap, because a self contained prompt does not need the previous conversation in order to work.
If I do need a second prompt, what goes in it?
The material. Not a reference to the material.
32 of our 455 prompts exist to continue work that already happened, and all 32 carry an inline paste slot. Not one of them says "use the draft above" or "take what you just wrote". They are spread across 13 of the 15 packs, which tells you this is a house habit rather than one author's tic: 6 in Job Seeker, 6 in Health and Wellness, 4 in Educator, 3 each in Ecommerce and Freelancer Toolkit, and 1 or 2 in each of the rest.
The two clearest cases are prompts whose entire purpose is to be step two of something. Our Twitter Thread to Long Form Blog Post prompt takes a thread that already exists and asks you to paste it in full, plus the engagement numbers, plus the top replies, plus your theory of why it worked. Our Content Repurposing Map takes a finished blog post and asks you to paste the content or a summary of it. In both cases the source is usually still on screen from the previous message, and both prompts ask for it again anyway.
That habit buys three things. The prompt works in a fresh chat. It works next week, when the conversation is gone. And it does not slowly degrade as the earlier exchange gets pushed further back in a long thread. If you are going to split, split cleanly, and give the second prompt everything it needs as though the first had never happened.
What about the one prompt that does wait?
It is a mock interview, and it is instructive precisely because it is the exception.
Act as a tough but fair interviewer for this exact role at this company. Ask me 10 questions, one at a time. After I answer each question: 1. Rate my answer (Strong / Acceptable / Needs Work) 2. Tell me specifically what was good 3. Tell me what was missing or could be stronger 4. If "Needs Work", rewrite my answer to show what a great response would sound like Start with question 1 now. Wait for my response before moving to question 2.
Here the exchange is the deliverable. You are practising being asked things, so a single block of ten questions with ten model answers would be useless. But look at what the prompt still does in that one opening message. It sets the role, the company, the industry, the interview type, the rating scale, the per answer format and the stopping rule, all before the first question is asked.
So even the genuinely conversational case is not "improvise a conversation". It is one prompt that designs the conversation, and then a conversation that runs inside it. The single turn discipline did not disappear. It moved up a level.
A prompt you can copy
This one is for when you already have four half formed asks in your head and you are about to start typing them one at a time. Give it the mess and it will hand back one prompt, or tell you honestly that you have two jobs.
I am about to ask an AI assistant for help with something, and I am not sure whether it is one request or several. Here is everything I want, in whatever order it came out of my head: [DUMP EVERYTHING YOU WANT. Bullet points, half sentences, contradictions, all of it. Do not tidy it up first.] Here is the material I already have: [PASTE ANY DRAFT, DATA, BRIEF, EMAIL OR DOCUMENT THIS IS ABOUT, OR WRITE "none".] Before writing any of it, do this: 1. Count the jobs. Say how many genuinely separate subjects are in the dump above. Two things are the same job if changing one would force a change in the other. If they are separate, say so and treat them separately. 2. For each job, list what I actually want back from it as a numbered list of deliverables. Name each one as a thing, not as an activity. 3. Flag the dependencies. Say which deliverables have to be written with another one in view, for example a summary of something else in the same answer, or a comparison between two of the items. 4. Flag what is missing. List the facts you would need from me that the dump does not contain, as bracketed slots I can fill in. 5. Write the prompt. One self contained prompt per job, each carrying the context, the bracketed slots and the numbered deliverables. Each prompt must work as the first message in an empty conversation, so do not refer to anything outside it. 6. Say what to leave out. If anything in my dump is better handled by me rather than by an assistant, say which and why. If the honest answer is that this is one job, say so and give me one prompt. Do not split a single job into a sequence of steps.
Step 1 is the one that earns its keep. Most of the time the answer is "this is one job and you were about to make it four", and occasionally it is "these are two jobs and you were about to jam them together", which is the other failure and about as common.
How this was measured, and what it does not show
The figures are direct counts over the 455 prompts published in our 15 packs, taken on 15 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 single turn figure was produced by testing all 455 prompt bodies against eleven separate patterns for conversational structure: waiting for a reply, one at a time, asking the user questions before answering, refusing to proceed until confirmation, referring to what the user will say next, naming a second or follow up prompt, pausing, and stopping to check. Four prompts matched at least one pattern, and all four were then read in full. Three were false positives, and they are worth naming so the number is checkable. One matched "pause" because it inserts jump cut markers where the creator can pause filming. One matched because a deliverable it produces is a schedule of "ask me anything" sessions for a new hire. One matched "I'll paste" in a sentence that introduces an inline paste inside the same prompt. Only the mock interview quoted above was genuine, so the count is 1 of 455, not 4.
Deliverables were counted as top level numbered lines inside the prompt body, which is why the median across all 455 is 4 while the median among the 276 prompts that actually use a numbered list is 5. Input slots were counted as bracketed placeholders. The count of prompts naming two or more kinds of output was restricted to the numbered deliverable lines, so a noun that appears only in a constraint does not inflate it. Paste prompts were matched on an inline paste instruction and every match was checked by eye.
The limit is the usual one and it is real. This is a census of prompts we believe work, not a controlled test of one prompt against four. We did not run that test and nothing above should be read as though we did. The mechanism in the second section, about later messages being written in the shadow of earlier ones, is reasoning, and it is labelled as such where it appears. What the census supports is narrower and still worth knowing: 455 prompts written for paying users across 15 professions, by people whose incentive is that the prompt works on the first try, are single turn 454 times out of 455. 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 splitting up a prompt
Put everything about one job in one prompt, and start a new prompt when the job changes rather than when the list gets long. Across the 455 prompts we publish, 454 are written to be answered in a single turn. The median one hands over 7 facts and asks for 4 separate things back, and 91 of the 455 ask for two or more different kinds of output at once, such as a diagnosis plus a rewrite plus a changelog. The thing that earns a second prompt is a second subject, not a longer list.
Because each message after the first is answered in the shadow of the one before it. When you ask for the outline, then the intro, then the conclusion, the intro is written without knowing how the piece ends and the conclusion is written to match an intro the model has already committed to. Constraints you mention in message four cannot reach back and change message two, so you either accept the mismatch or ask for a rewrite. A single prompt that names all four parts lets the model plan across them before it writes any of them. That is reasoning about how the pieces fit rather than a measured effect, and we say so plainly: what we measured is that 454 of our 455 prompts are built the single turn way.
Three situations, and none of them is length. First, the subject changes: the customer emails and the quarterly budget are two jobs and belong in two prompts. Second, you need to react before you can specify: you genuinely cannot write the brief for step two until you have read step one, which is what our 32 teardown and rewrite prompts assume. Third, the input is bulky enough that mixing it with a second job makes the instruction hard to find. Notice that all three are about the work, not about the size of your request.
Paste the material back in. 32 of our 455 prompts exist specifically to continue earlier work, and every one of them carries an inline paste slot rather than saying use what you wrote above. Our Twitter Thread to Long Form Blog Post prompt asks you to paste the full thread even though the thread is usually still on screen, and our Content Repurposing Map asks you to paste the source content or a summary of it. That is a deliberate habit: a prompt that carries its own inputs works in a fresh chat, works next week, and does not quietly degrade as the conversation gets long.
No, and conflating the two is what pushes people to split. The median prompt in our library is 174 words, which is about a minute of typing, and the longest is 424. What makes those prompts feel substantial is not difficulty but bookkeeping: 3,239 bracketed slots across the library, a median of 7 per prompt, each one a fact you were going to have to supply anyway. You are not making the task harder by writing it all down. You are moving the work from four rounds of correction into one round of specification.
Start a new chat whenever the subject changes, because a self contained prompt does not need the history. Every prompt we publish is written to work as the first message in an empty conversation, which is why they all state the role, the context and the output shape rather than pointing back at anything. The practical benefit is that a stale earlier answer cannot influence a new one. If you are continuing genuinely related work, staying put is fine, but paste the material you are building on regardless.
Then the back and forth is the deliverable and you should say so explicitly in the first prompt. Exactly one of our 455 prompts does this: a mock interview that instructs the model to ask 10 questions one at a time, rate each answer, and wait for the reply before moving on. Note what it still does in that single opening message. It sets the role, the company, the interview type, the rating scale and the stopping rule up front. Even the conversational case is one prompt that designs the conversation, not a conversation you improvise.
Related reading and next steps: if the worry behind splitting was that one prompt would get too long, how long should a ChatGPT prompt be measures every prompt in the library and settles it. For what has to be inside that single prompt, read what information should I give ChatGPT, and for how to make its several deliverables come back in the shape you asked for, how to get ChatGPT to follow the format you asked for. If your instinct to split came from answers that all sounded the same, the cause is in why ChatGPT gives generic answers, and the related habit of asking for many versions instead of one good one is measured in how many options should I ask ChatGPT for. To start from prompts that are already self contained, browse the prompt packs or read the how to use guide.