4d • General Discussion
Stop Optimising Your Prompts. Start Improving Your Inputs
Let's talk about something that I think a lot of people get backwards when it comes to prompt engineering.
We're constantly trying to get more out of AI.
Better answers. Better reasoning. Better formatting. More detail. More consistency. More creativity.
And what do we do?
We add another instruction. Another constraint. Another paragraph explaining what we don't want. Another rule to stop the AI doing something it did three conversations ago.
Before long, we've created a prompt so complicated that maintaining it becomes a job in itself.
And here's the question I want you to consider.
What if the problem isn't that you're asking too much of AI? What if the problem is that you're not giving it enough of what it actually needs?
There's a difference.
Think about hiring somebody to do a job. You could give them a twenty-page instruction manual explaining every possible thing they might encounter.
Or you could explain what you're trying to achieve, give them the information they need, establish what a good result looks like, and let them get on with it.
That's the distinction we're exploring today.
Because good prompt engineering isn't about squeezing every possible capability out of a language model. It's about designing an interaction where the model has enough information to produce something useful.
And that brings us to the architecture of a standing prompt.
A standing prompt is essentially the foundation you establish before the individual task begins. It's the persistent instruction set that tells an AI how to work with you.
But here's where people get carried away.
They start trying to account for absolutely everything.
Every possible scenario. Every possible mistake. Every possible formatting preference. Every possible response they might dislike.
And the prompt grows.
Then it grows again.
Eventually, you've got a collection of instructions that contradict one another, overlap, or compete for attention.
This is over-optimisation.
And it's particularly interesting because the intention is usually good. You're trying to make the system more reliable.
But adding more instructions doesn't automatically make a prompt more reliable.
Sometimes it makes the actual objective harder to recognise.
So, what should you give AI instead?
Start with the outcome.
What are you actually trying to achieve?
Not what should the AI say, not which framework should it use, not how many steps should it follow.
What does success look like?
Let's say you're creating a prompt that turns rough ideas into finished social media posts.
You could write thirty instructions about tone, sentence structure, formatting, engagement techniques and vocabulary.
Or you could explain that you want rough thoughts transformed into clear, natural social media posts that preserve your original meaning, sound human, and are ready to publish.
Now you've established a destination.
Next comes context.
And this is where I think prompt engineering becomes considerably more interesting.
Because context is not just additional information. It's information that changes the quality of the decision.
Who is the audience? What do they already know? What's the purpose of the content? What information must be preserved? What would make the result inappropriate?
You're not giving AI more information for the sake of it.
You're giving it the information that makes the difference.
And there's another important layer, boundaries.
What should the AI avoid changing? Where should it exercise judgement? When should it stop and ask you a question rather than making an assumption?
These boundaries are particularly useful in standing prompts because they establish how the system should behave across different tasks.
Then you have the output.
What do you actually want back?
A finished document? A short explanation? A table? Three alternatives? Something you can copy directly into another application?
Make that clear.
And finally, give the AI a way to recognise when it's finished.
This is something people often overlook.
You don't necessarily want an AI that keeps improving something indefinitely. You want one that understands what a satisfactory result looks like and knows when further changes aren't adding meaningful value.
Now, let's return to the bigger problem.
You will never get one hundred per cent quality, one hundred per cent of the time.
Language models are probabilistic systems. Their outputs can vary, even when the instructions remain unchanged.
And different models interpret instructions differently.
So, the goal isn't to build a magical prompt that never produces a disappointing response.
The goal is to improve the consistency of useful responses, reduce unnecessary corrections, and make the system easier to work with.
That's a much more realistic target.
There's also a misconception worth challenging here.
People sometimes assume that the longest, most sophisticated prompt must be the best prompt.
It doesn't.
A simple prompt can be excellent when the task is simple. A complex prompt can be necessary when the task has multiple stages, strict requirements or important consequences.
The architecture should match the problem.
Not the other way around.
And this is where I want to leave you with a practical exercise.
The next time you build a standing prompt, don't start by asking yourself what else you can tell the AI to do.
Ask yourself what information it is currently missing.
What does it need to understand about the objective?
What context would change its answer?
What boundaries genuinely matter?
What would a good result look like?
And perhaps most importantly, what can you remove without making the result worse?
That last question is incredibly valuable.
Because prompt engineering isn't just the art of adding instructions. It's also the discipline of knowing which instructions don't need to exist.
Think of your prompt as an operating environment, not a cage.
Give the AI a clear destination, the right information, sensible boundaries and enough room to do the work.
Then test it.
Look at where it succeeds, where it struggles and what information would have helped.
Make the smallest meaningful adjustment.
Don't rebuild the entire thing every time you get an imperfect answer.
And don't confuse a more complicated prompt with a better engineered one.
Ultimately, the question isn't:
How much can I get out of AI?
It's:
What do I need to give AI so that it can give me what I actually want?
That small change in perspective can completely change the way you approach prompt engineering.
And it might save you an awful lot of time spent rewriting instructions that weren't the real problem in the first place.
That's where I'd start.
23
7 comments
Stop Optimising Your Prompts. Start Improving Your Inputs
skool.com/the-ai-advantage
Founded by Tony Robbins, Dean Graziosi & Igor Pogany - AI Advantage is your go-to hub to simplify AI and confidently unlock real & repeatable results
Leaderboard (30-day)
1
+1677
2
+1180
3
+685
4
+539
5
+491
Powered by