What it is
A deliverable is the thing you will copy, send, paste or file when the model is done. Stating it takes a verb the model can carry out and a noun it can produce: draft a reply, rank these vendors, write the three sentences that open the board report. Requests such as “look at this”, “help with” or “thoughts on” name an activity instead, and the model has to guess what you want out of it.
Anthropic’s prompting guide suggests a test for this. Show the prompt to a colleague who has little background on the task and ask whether they could follow it. A colleague handed “thoughts on this email?” would ask what kind of thoughts. A model does not ask. It picks one kind and gives it to you with confidence.
In the thinkhard.si study of 319 one-line work requests, a one-line request passed this check about two times in three, more often than any other. The third time is the one that costs you a second round.
When it helps
It helps whenever a request could produce more than one kind of answer. “Review this contract” could mean a summary, a list of risks, a marked-up copy or a view on whether to sign. Each is a reasonable reading, and only one is what you need before Thursday’s call. Name that one.
It also helps when the answer is going somewhere specific, such as an email, a report or a spreadsheet row. Say what the thing is, and the model produces the piece you will use instead of writing about the task.
Watch for the soft verbs: look at, help with, check, improve, go over, thoughts on. Each is fine as an opening, as long as a deliverable follows it. “Check this invoice” becomes “check this invoice and list every line that does not match the purchase order.”
When it does not
When you do not yet know what you want, naming a deliverable commits you to a shape too early. If you are still working out the problem, ask for options, or have the model ask you questions first, and name the deliverable once you know it.
A deliverable is not a brief. “Write the project update” names the thing and gives the model nothing to build it from, so you get a generic update with progress it made up. The deliverable says what to make; the facts and the reader say what goes in it.
The model makes what you name, even when something else would serve better: ask for an email and you get one, even if a two-line chat message would do. And a deliverable that is really several jobs, such as “write our Q3 strategy”, comes back thin everywhere. Split it into pieces you could each check.
Before and after
Three requests from different jobs. Each score comes from the Prompt Grader’s free ten-check estimate, written for any chat model; paste either version in and you will get the same number.
Thoughts on this email from a customer: """ Hi, I was charged twice for my March invoice (#4471). I need the duplicate refunded before my card statement closes on the 28th. Who do I talk to? """
Draft a reply to the customer below, from our billing team, that confirms the duplicate charge and tells them when the refund will reach their card. """ Hi, I was charged twice for my March invoice (#4471). I need the duplicate refunded before my card statement closes on the 28th. Who do I talk to? """
The first version gets an opinion about the email, which might be a tone check, a summary or advice on how to respond. The second gets the reply itself, addressed to the customer and answering the question they asked about the refund date.
Help with the Q3 numbers below. Revenue: $2.41M against a target of $2.60M New customers: 38 against a target of 45 Churn: 2.1%, down from 3.4% in Q2
Write the three sentences that open our Q3 board report: whether we hit our targets, by how much, and the main reason. Revenue: $2.41M against a target of $2.60M New customers: 38 against a target of 45 Churn: 2.1%, down from 3.4% in Q2
Asked for help, a model may comment on every number in turn. Asked for three opening sentences, it has to choose what matters and say it in the order a board reads it, which leaves you with text you can paste rather than notes to rewrite.
Improve this job post. [the job post, pasted in full]
Rewrite this job post as a 250-word version for LinkedIn. Keep the salary range and the remote policy word for word. After the post, list what you cut. [the job post, pasted in full]
“Improve” lets the model decide what better means, and that can mean longer. The second version names the thing to produce and where it will run, and the list of cuts shows you what was lost before anyone applies.
None of the after versions is finished. Each now passes Task; the other patterns in this group add the rest.
The line to add
Put this first, before any material, and fill in the blanks:
{{verb}} {{the_thing_you_want_back}} that I can {{send, paste or file}} as it is.
Filled in, it reads like “Draft a reply to this customer that I can send as it is” or “Write the summary row for this week’s tracker that I can paste as it is.” Then read it back and ask: if the model did exactly this, would I have something I could use without asking again? If the answer is “sort of”, the deliverable is still vague.
In ChatGPT, Claude, Gemini and Copilot
This is the pattern that changes least between assistants, and each vendor’s guidance starts with it. Anthropic’s guide for Claude opens with being clear and direct, and says being specific about the output you want can improve results. Google’s guide for Gemini opens with clear and specific instructions, and Google’s prompting guide for Gemini in Workspace makes the task one of the four areas of a prompt.
Microsoft’s guide to writing Copilot prompts lists four parts, goal, context, expectations and source, and says a clear goal is the only one required. OpenAI’s guide to its latest models, the ones behind ChatGPT, says the model can often infer the steps from context, so you need not prescribe each one, but you should still give the success criteria and constraints. In every case the deliverable is the outcome you state.
The sentence itself needs no rewording between assistants; the layout differences, covered in the guide to Claude, GPT and Gemini, are about everything around it.
Key takeaways
- Name the thing you want back, with a verb the model can carry out.
- “Look at”, “help with” and “thoughts on” name an activity. Follow them with a result.
- The model makes what you name, so name what you will actually use.
- A deliverable is the start of a brief, not all of it: the facts and the reader come next.
- If you do not know what you want yet, ask for options or questions first.
Questions
What is the most important part of a prompt?
Saying exactly what you want back. Every other part, from the facts to the length, shapes a result, and without a named result the model has to guess what it is shaping. After that, the facts the model cannot know matter most. The ten checks cover the rest in order.
Why does ChatGPT give me a different kind of answer than I wanted?
Usually because the request named an activity instead of a result. “Help me with this report” could mean edit it, summarize it or critique it, and the model picks one. Name the result, such as “list the three weakest sections and say why each is weak”, and the answer changes shape to match.
How do I write a clear task for AI?
Use one verb the model can carry out and one thing it can produce: draft a reply, rank these options, write three sentences for the report. Then read it as a new colleague would and ask whether they could start without checking what you meant. The free Prompt Grader marks the Task check as you type.