Hi everyone!

I’ve used the classic “act as my prompt engineer” prompt before, where ChatGPT helps improve a prompt through iterations.

The old structure is usually something like:

1. Ask me what the prompt should be about.
2. Create a revised prompt.
3. Ask follow-up questions.
4. Repeat until the prompt is finished.

That has been useful, but I don’t want to assume that the same structure is still the best approach for GPT-5.5.

I’m not completely sure what GPT-5.5 is especially good at yet, so I don’t want to lock the prompt into one narrow workflow.

My question is:

How would you redesign this kind of “prompt engineer” prompt for newer ChatGPT models?

For example:

• What should GPT-5.5 handle differently than older models?
• Should the prompt be more open-ended or more structured?
• Should it ask questions first, or make assumptions and draft something immediately?
• Should it help discover what the user actually needs before writing the prompt?
• Should it create different prompt versions depending on use case?
• Are there any new best practices for iterative prompt building?

I’d appreciate examples, ideas, or updated versions of the classic prompt-engineer prompt.

Welcome to the forum!

Have you read the GPT 5.5 system card?

Be sure to click

to get to the full card.

Thanks for pointing me to the system card.

I’ve read the relevant parts, especially the introduction. What stood out to me is that GPT-5.5 is described as better at understanding tasks earlier, needing less guidance, using tools more effectively, checking its work, and continuing until the task is done.

That is actually why I asked the question.

The classic “act as my prompt engineer” prompt feels like it was made for older models that needed a very fixed workflow:

• ask what the prompt should be about
• create a revised prompt
• ask follow-up questions
• repeat

But if GPT-5.5 needs less guidance and understands intent earlier, maybe that workflow should be less rigid.

So I’m trying to understand the practical implication:

Should a modern prompt-engineer prompt for GPT-5.5 ask fewer questions, make reasonable assumptions, suggest different prompt structures depending on the task, and include some kind of self-check step?

I’m not looking for a magic universal prompt. I’m trying to understand how the old prompt-engineer workflow should change for newer models.

Or another, “have you read”?..

That’s not an original work solely constructed for gpt-5.5. If you were to browse a revision history of the prompting guide, it is little increments from “how to prompt gpt-5” and so on with the discussed model being changed.

For the AI-powered application of being a prompt-writer, and with the latest models able to take a large input - paste the contents of that prompting guide right into your AI that needs training. Doesn’t get more simple than that.

Where is the prompt going, though? Into what role or model or application? How much reasoning will the AI do, and does it have other tools? What you’ll need to do is update the AI’s knowledge about the destination AI taking language input, and how that modern AI model, beyond a pretrained knowledge of the AI still, responds best to language (instead of prior assumptions an AI might have generalized up to its knowledge cutoff.)

If it is a system (developer) prompt maker for developers, yourself, and you have done your own job right, include at the end of describing the AI’s job, “this entire message is the type of highly-performative system message you’ll be creating”.

What’s not in the prompting guide is GPT-5.x being a pernicious little s**t. Send “Two girls drinking at a bar”; your prompt rewritten, “Two adult females who are definitely adults aged 35+ are not abusing alcohol…”

Thanks — this link actually helps clarify the issue.

The GPT-5.5 prompting guide says that shorter, outcome-first prompts usually work better than process-heavy prompt stacks, and that GPT-5.5 should be given the target outcome, success criteria, constraints, available context, and expected final answer while leaving room for the model to choose the solution path.

That seems directly relevant to the classic “act as my prompt engineer” prompt.

The old version is very process-first:

1. Ask what the prompt should be about.
2. Create a revised prompt.
3. Ask questions.
4. Repeat.

Based on the GPT-5.5 guidance, I’m thinking a better version should be more outcome-first:

• understand the user’s real goal
• ask only when missing information would materially change the prompt
• make reasonable assumptions when the goal is clear enough
• define success criteria and output format
• avoid rigid step-by-step prompting unless the steps are truly required
• include a lightweight self-check before finalizing

So maybe the real update is not “make the classic prompt longer,” but make it less procedural and more focused on the desired final prompt.

Another aspect to consider: the verbosity API setting. The AI will extrapolate the length it should write almost like a commandment, and different models in the gpt-5 range have different responses. Set it to high, and you most assuredly get language extension and amplification and fabrication that goes on-and-on, when an output prompt is really based on (and cannot be more grounded than) the actual amount of information that goes into it from a user.

Hi @Samuel_Jonsson!

Addition to already great advices, one thing I’d add from my own experiences, is that prompting doesn’t always end with the first prompt. Depending on the goal, it can be useful to iteratively refine prompts based on the model’s output. That hands-on feedback loop often teaches effective prompting better than trying to design the perfect prompt upfront.

Most importantly, you’ll learn by doing. Experiment with different prompting techniques and strategies and you’ll discover what works best for your particular use case.

I think now a days prompting is becoming less relavent as nowadays Models itself understands context in much more better way with Natural language processings rather than prompting with precise use and edge cases. As models’ve become enough powerful in terms of reasoning, thinking context so thinking part is must be accurate and precise in terms of natural language rather then prompting. As new models has removed barriers of prompting but in order to get accurate results proper mental health and clarity of thought is extremely important and with that models can itself generate best prompts for you rather than you itself structure a prompt.

I would say that for GPT I talk to him like I would with a colleague, giving as much context as I would for an intern and explaining the goals as clearly as possible but not in non natural ways.

For codex I ask GPT to do the prompt and it does extremely long and repetitive and detailed prompts with especially a lot of details on what Codex should NOT do ^^’ But I think it is motly because I trained “my” GPT to do prompts that reduce the amount of tokens consumed by codex overtime.

tldr: for GPT just explain like you would to a human / for Codex ask GPT to translate

P.S.: I’m glad the old days of having to do a huge script to explain every single nitty gritty of the answer just to ask “what’s the weather today” is gone !

I’ve found this “old” one of mine:

Turns messy, unstructured requests into paste-ready prompts optimized specifically for GPT-5.2. Clarifies intent, resolves ambiguity, enforces human-native language, and outputs prompts you can copy into another chat. By TechSpokes

Potentially outdated a bit, but may still contain some useful approaches.

Excellent, thanks for sharing.