Eleven prompts, nine named roles, one shape. AI Central's Pro Prompts for ChatGPT, drawn from OpenAI's Academy, is not a collection of clever wording. Every prompt in it states a role, opens a bracket for context the model cannot know, and closes by specifying the format of the answer. Ten of the eleven name that output shape outright. Copy the structure rather than the sentences and the pack keeps working long after the examples date.
The shape underneath all eleven
AI Central assembled the pack as a ready to use set for work, crediting the prompts on every page to OpenAI's Academy. Read the eleven back to back and the striking thing is not the range of jobs they cover, it is how little the structure changes between them.
Three parts repeat almost without exception. A role or situation is stated up front. At least one bracketed slot waits for context the model has no way to know. A closing clause describes the artefact you want back.
Ten of the eleven end by naming that output explicitly. Counted one by one, the pattern is hard to miss.
- The sales prompt asks for email ready text.
- Marketing asks for a short briefing with bullet point insights.
- The manager prompt asks for three objectives with three or four key results each.
- Finance asks for a table carrying metrics, source links and a short analysis.
- HR asks for a description, an expected impact and an effort level per idea.
- Product asks for a comparison table plus recommendations.
- Executives ask for three guiding principles and a simple comms calendar.
- Customer success asks for a week by week table with task owners and goals.
- The rewrite prompt asks for a tone that is clear, respectful and concise.
- The agenda prompt asks for sections with time estimates and a goal each.
Only the engineering prompt leaves the shape open, and even that one names its two deliverables, the most likely causes and the next steps for mitigation.
That consistency is the technique. The sentences are replaceable. The scaffold is not.
The brackets carry the weight
Every prompt in AI Central's pack contains at least one bracketed slot, and the sales entry contains three.
Write a short, compelling cold email to a [job title] at [company name] introducing our product. Use the background below to customize it. Background: [insert value props or ICP info]. Format it in email-ready text.
That is the sales prompt exactly as AI Central publishes it. Look at what the three brackets are doing. Two of them identify who is being written to. The third, the value propositions or ideal customer profile, is the only place the model can learn what you actually sell.
Paste that with the brackets empty and you get a fluent template about nothing. This is the most common failure with borrowed prompts, and it is not the prompt's fault. The wording is a wrapper. The bracketed context is the payload, and the payload is the part only you can supply.
Read the whole pack that way and it becomes a specification of what each job requires you to bring. Quarterly goals need business context, company objectives and recent performance. A production debug needs logs, metrics and recent changes. A customer onboarding template needs the customer type. Competitor research needs the target product. None of that is guessable.
Naming the format is the step most people skip
The closing clause in each of AI Central's prompts does quiet work. It converts an open question into a deliverable with edges.
The HR entry is the clearest case. It asks for five ideas, then dictates exactly how each one arrives.
Present each idea with a short description, expected impact, and implementation effort level
That single clause, from AI Central's HR prompt, turns a brainstorm into something you can take to a decision. Ideas with an effort level attached can be sorted. Ideas without one arrive as a list you still have to triage by hand.
The same prompt sets its constraints before it asks for anything at all.
Consider our hybrid work model, current engagement scores, and time/resource constraints.
Constraints stated up front, as AI Central does here, do more for output quality than any adjective. Practical is a hope. A hybrid work model, a current engagement score and a resource ceiling are filters, and a model can actually apply a filter.
Three of these are research jobs, not writing jobs
The marketing, finance and product entries all open with the verb research, and that changes what you are responsible for once the answer lands.
Two of the three demand receipts. The marketing prompt tells the model to cite sources. The finance prompt goes further.
Provide a table with metrics, source links, and a short analysis of how we compare
AI Central's finance entry asks for expense ratio benchmarks across five comparable companies with links attached. That is the right instinct. A benchmarking answer without sources is not a shortcut, it is a liability you have to redo before anyone senior sees it.
So treat the research prompts differently from the writing ones. On a rewrite you judge the output by reading it. On a benchmark you judge it by opening the links. Same tool, different burden of proof, and the pack is explicit about which is which.
Bounded asks beat open ones
Five of the eleven fix a number before asking for anything. Three objectives with three or four key results each. Five comparable companies. Five engagement ideas. Three competitor onboarding flows. Three guiding principles for internal communications.
A bound forces a ranking. Asked for ideas, a model produces as many as it can and puts no weight behind any of them. Asked for five, it has to choose, and the choosing is where the value sits. The numbers in AI Central's pack are small on purpose.
How to actually use this
If you are starting out, take the one prompt that matches something you already do every week and fill in every bracket properly, once. Not a phrase per slot, the real context. Then save the filled version, because the copy with your context in it is the asset. The blank never was.
If you already prompt daily, lift the closing clause rather than the whole prompt. An instruction to format as a week by week table with task owners and goals is transferable to a dozen tasks that have nothing to do with customer onboarding. Build a short library of output shapes you trust and attach them to whatever you are asking.
If you run a team, standardise the brackets, not the wording. AI Central's pack is effectively a list of what context each function owes the model, and a team that agrees on that list gets consistent output without anyone memorising a prompt.
AI Central closes with the least complicated advice in the whole pack.
Don't wait for inspiration
The instruction AI Central leaves you with is to drop a prompt, polish the output, and take the result. These prompts reward a first draft and a second pass, not deliberation.
What am I supposed to put in the square brackets?
Whatever the model cannot know about your situation. Across AI Central's pack the slots ask for a job title, a company or sector, a product, a topic, a meeting length and its attendees, or a block of raw context such as logs, metrics, recent performance or your value propositions. If a bracket is hard to fill, that is a sign the task is not defined yet, not a reason to leave it empty.
Do I have to copy the wording exactly?
No. The wording is the least important part. What carries across is the pattern AI Central repeats in all eleven, state the role, supply the context, name the format you want back. Rewrite the sentences in your own voice and the results hold.
Which one should I start with?
The one attached to a task you repeat. The rewrite prompt and the meeting agenda prompt are the lowest friction, because they need almost no setup, just a block of text, or a topic, a duration and a list of attendees. The finance and product research prompts pay off more, but they cost real context up front and real verification afterwards.
Why insist on a table or a bullet list instead of just letting it answer?
Because the format decides whether you can use the answer. Ten of the eleven prompts in AI Central's pack name an output shape, and each shape maps to a next step. A table gets compared. A week by week plan gets assigned. An unstructured paragraph gets read once and forgotten.