Latest / How I Work / Your prompts are stuck in 2025. Here are 5 fixes
Transcript
- 0:00Speaker 1: Tell me if you can relate to this. You've got
- 0:03Speaker 1: a prompt that you've been reusing for a year. It
- 0:06Speaker 1: opens with, you are a world-class expert. It asks the
- 0:11Speaker 1: AI to think through its answer, and it has a
- 0:13Speaker 1: couple of rules in capital letters, so it knows you
- 0:17Speaker 1: mean business. But lately, though, maybe those results have felt
- 0:21Speaker 1: a little bit off. And the reason why is because
- 0:25Speaker 1: the models have moved on and a lot of the
- 0:28Speaker 1: habits that we picked up in 2024 and 2025 now
- 0:32Speaker 1: actually get in the way. Today's AI takes your words
- 0:37Speaker 1: very literally and it already does the thinking that you
- 0:41Speaker 1: used to have to ask it to do. In today's episode,
- 0:45Speaker 1: Neo and I go through five ways that prompting has changed.
- 0:50Speaker 1: You're going to learn why. what a persona really changes
- 0:53Speaker 1: in a response, why a goal plus a finished example
- 0:57Speaker 1: often beats spelling out every step. And we're also going
- 1:01Speaker 1: to cover the importance of a clear checklist to get
- 1:05Speaker 1: AI to actually check its work before handing anything over.
- 1:08Speaker 1: By the end of today's show, you are going to
- 1:12Speaker 1: have prompts that are much shorter and clearer and also
- 1:16Speaker 1: much more likely to get it right the first time.
- 1:24Speaker 1: Welcome to How I AI with me, Dr. Amantha Imber
- 1:27Speaker 1: and Neo Applin, head of Inventium AI. Each episode, we
- 1:32Speaker 1: share one practical way to use AI better at work
- 1:35Speaker 1: and in life. No fluff, no tech jargon, just things
- 1:39Speaker 1: you can use straight away. In our training programs at
- 1:44Speaker 1: Inventium AI, we, of course, have always talked about prompting,
- 1:50Speaker 1: but the way that we talk about prompting has changed
- 1:53Speaker 1: a a lot since 2024 and has also changed since 2025. So, Neo,
- 2:00Speaker 1: I would love to know what are the biggest changes
- 2:05Speaker 1: that have happened in what good prompting looks like in
- 2:09Speaker 1: late 2026 as we are recording this in October versus, say,
- 2:12Speaker 1: a year ago? Where should we start?
- 2:20Speaker 2: My first one that I'd look at is probably persona.
- 2:23Speaker 2: It used to be that you would imbue this role
- 2:27Speaker 2: into the AI. You are a world-class expert in X, Y,
- 2:31Speaker 2: and Z. And then it would be able to give
- 2:33Speaker 2: you a better result than just, hey, tell me the thing.
- 2:37Speaker 2: And that certainly was the case back in 2024, a
- 2:39Speaker 2: little bit of 2025, and it's certainly gone out of place.
- 2:43Speaker 2: Where role or persona or perspective, which is what we
- 2:47Speaker 2: probably use is really valuable, is It then comes to
- 2:50Speaker 2: perspective or the language that comes out of it, but
- 2:52Speaker 2: it doesn't give you a better result. It just gives
- 2:55Speaker 2: you a different result. It's kind of like if you said,
- 2:58Speaker 2: explain it like a high school teacher versus explain it
- 3:00Speaker 2: like a university professor. It's not a better result. It's
- 3:04Speaker 2: just a different one that might help you in a
- 3:05Speaker 2: different way. So perspective is good. But, yes, certainly don't
- 3:09Speaker 2: say you're smarter than Einstein with a 2 million IQ.
- 3:14Speaker 2: It's going to make up stuff hilariously and certainly won't
- 3:17Speaker 2: actually improve anything. So, yes, give it a perspective rather
- 3:20Speaker 2: than imbue a role. Okay.
- 3:24Speaker 1: Tell me about micromanagement because previously, certainly in 2024 and 2025, we–
- 3:33Speaker 1: we would need to micromanage AI's response and how it
- 3:38Speaker 1: went about its job. But we don't need to do
- 3:41Speaker 1: that anymore, do we?
- 3:42Speaker 2: Well, we do and we don't. I think I disagree
- 3:45Speaker 2: a little bit with some of the advice that the
- 3:47Speaker 2: companies have. And I'll tell you why. The tools have
- 3:51Speaker 2: changed a bit so that they don't need to be
- 3:54Speaker 2: told necessarily how to do things every single time. And
- 3:58Speaker 2: they're relatively smart and they can go about and find
- 4:00Speaker 2: their things. And so what you can do is just say,
- 4:02Speaker 2: here's my goal and here's what good looks like. You
- 4:06Speaker 2: sort out how you get there. And they've found that
- 4:09Speaker 2: that is more effective because if you say, look, you
- 4:12Speaker 2: must do it my way, that may be less effective
- 4:14Speaker 2: for the AI to do it and then effectively then
- 4:17Speaker 2: hit against walls, rework, all of those kind of things.
- 4:20Speaker 2: And I do love this. The goal and what good
- 4:23Speaker 2: looks like, not just telling it, by the way, I
- 4:25Speaker 2: love showing it what good looks like. So if you've
- 4:27Speaker 2: got an example of that thing you've done before and say,
- 4:30Speaker 2: it's got to be like this, then understand. So it's
- 4:32Speaker 2: rather than pictures of a thousand words, a completed document
- 4:37Speaker 2: is a million words, if you'd like. So you can
- 4:39Speaker 2: get a much better outcome there. But sometimes... we do
- 4:42Speaker 2: need to do things in a particular way for our jobs.
- 4:46Speaker 2: You might have a meeting report template and it needs
- 4:51Speaker 2: to be done in a very particular way. And maybe
- 4:54Speaker 2: some of the sources need to come from very specific
- 4:56Speaker 2: areas or calculated in a particular way. So simply just going, hey,
- 5:00Speaker 2: write me a report, here's what it needs to look like,
- 5:02Speaker 2: doesn't necessarily give you the best results. So I think
- 5:06Speaker 2: it's a continuum. Sometimes it's very much like, just give
- 5:08Speaker 2: it the goal and what good looks like, Other times
- 5:10Speaker 2: you do need to say, do this, then this, then that.
- 5:14Speaker 2: But the ultimate thing is this, if you don't need
- 5:18Speaker 2: to be a micromanager, don't be a micromanager with the
- 5:21Speaker 2: AI because you don't need to be anymore. Okay.
- 5:24Speaker 1: Now, something that we used to recommend in our programs
- 5:28Speaker 1: when we were asking AI to do something that was
- 5:33Speaker 1: potentially quite a complex task is we would ask it
- 5:36Speaker 1: to Think through your answer. Can you talk about that
- 5:40Speaker 1: and why we don't actually need to do that anymore?
- 5:44Speaker 2: The early models don't think. Thinking is a really odd
- 5:51Speaker 2: word because you and I think in different ways than
- 5:54Speaker 2: the AI does think. Let me tell you about this.
- 5:57Speaker 2: The early models would predict words, and they still do now,
- 6:01Speaker 2: and it would give you the result and say, here's
- 6:04Speaker 2: what I think from my brain, or effectively just dump
- 6:06Speaker 2: out the result there. but what they found was by saying,
- 6:10Speaker 2: think through your answer, it would actually work out its
- 6:13Speaker 2: thinking process in its brain. It would actually, a weird
- 6:17Speaker 2: vomit on your screen, so you could actually see how
- 6:19Speaker 2: it was thinking. And that was before they created thinking models.
- 6:23Speaker 2: And so thinking models still do that kind of thinking,
- 6:25Speaker 2: but they just don't show you that boring text of,
- 6:28Speaker 2: I'm trying to do this, and I'm not quite sure,
- 6:29Speaker 2: is that right? And then this is the avenue I'm
- 6:31Speaker 2: going through. That kind of thinking out loud still happens,
- 6:35Speaker 2: you just don't see it on your screen. And so
- 6:38Speaker 2: if you're saying think through your answer with something which
- 6:41Speaker 2: is already doing a thinking behind the scenes, then you're
- 6:44Speaker 2: actually getting to perform thinking, even though it's already done
- 6:47Speaker 2: the job. And sometimes you get a worse response there.
- 6:50Speaker 2: So pretty much if you're on one of the tiny models,
- 6:53Speaker 2: maybe like a quick response you might have on it
- 6:57Speaker 2: or like a haiku, maybe you'll get a better response
- 7:00Speaker 2: with think through your answer, but you're much better off
- 7:02Speaker 2: just using a standard model that does the thinking these days.
- 7:06Speaker 2: So Don't get to perform the thinking if it's already
- 7:09Speaker 2: done the thinking.
- 7:11Speaker 1: Now, something that used to be helpful in 2024 and
- 7:15Speaker 1: perhaps 2025 as well is yelling at your AI in
- 7:17Speaker 1: all caps to make sure that it has heard you.
- 7:19Speaker 1: Tell me, do we still need to yell at our
- 7:21Speaker 1: AIs in late 2026?
- 7:33Speaker 2: If you want to, go right ahead. But what they've
- 7:38Speaker 2: done is, I think a couple of people were doing
- 7:40Speaker 2: things in different ways. So some people were just basically
- 7:42Speaker 2: yelling the whole time, thinking that being Donald Trump's on
- 7:45Speaker 2: all caps for your internal conversation would give you a
- 7:47Speaker 2: better result. That certainly was not the case. The emphasis
- 7:52Speaker 2: that yelling used to give you, always or never kind
- 7:56Speaker 2: of thing, was really good in that AI could tell, wow,
- 7:59Speaker 2: that's really important. Amantha doesn't usually... You'll use all caps.
- 8:02Speaker 2: Maybe that's pretty important. So I'll make sure that I
- 8:04Speaker 2: do that. So showing emphasis absolutely is still required and
- 8:09Speaker 2: still very good. What you might find these days is
- 8:14Speaker 2: if you use the words never or always or must,
- 8:17Speaker 2: it will take it as a absolute rule. And that
- 8:23Speaker 2: is a problem because occasionally it might go, well, it's
- 8:26Speaker 2: a better way to do it or a different way
- 8:28Speaker 2: to do it. And so I can kind of get
- 8:30Speaker 2: to the same goal without being told a particular process
- 8:33Speaker 2: way of doing things. So I would say only use
- 8:37Speaker 2: absolutes when you really need to use those absolutes, always
- 8:40Speaker 2: or never, and things like that. They've also said that
- 8:44Speaker 2: even though that does happen, occasionally what they've also said
- 8:47Speaker 2: is occasionally you actually get a better response, not by
- 8:50Speaker 2: being absolute, but saying, don't do this and here's why.
- 8:54Speaker 2: It does a much better job if you say why
- 8:56Speaker 2: it's necessary important or why it shouldn't be that output.
- 9:01Speaker 2: So you might say, needs to be 150 characters or
- 9:05Speaker 2: less because this is only a title that people are
- 9:08Speaker 2: going to skim through and I'm going to have 30,000
- 9:10Speaker 2: titles on the screen that people need to scroll through
- 9:12Speaker 2: to find the title. And it goes, oh, I understand
- 9:14Speaker 2: it's quick readability. You'll find you get much better adherence
- 9:17Speaker 2: there than an absolute rule. Okay.
- 9:21Speaker 1: We've got one final tip to give people about how
- 9:25Speaker 1: to do prompting better at time of recording in October, 2026.
- 9:31Speaker 1: And that relates to asking it to check things. Can
- 9:36Speaker 1: you explain a little bit about this?
- 9:39Speaker 2: It's part of that. It's the dovetail of giving it
- 9:41Speaker 2: a goal and what good looks like is saying, here's
- 9:44Speaker 2: a checklist that I'd like you to check off. So
- 9:47Speaker 2: before you give me that work, then please make sure
- 9:50Speaker 2: you've done all these things because AI now can work
- 9:53Speaker 2: longer than it used to before. Certainly if you've got say,
- 9:57Speaker 2: co-work or work, and now they're obviously putting these two
- 10:00Speaker 2: kind of features together, but it can work longer on things.
- 10:03Speaker 2: And what it can do is think, is this ready
- 10:06Speaker 2: for me to deliver to my friendly human? And you
- 10:08Speaker 2: can help it to do that by saying, look, here's
- 10:09Speaker 2: a checklist. So does it have this? Does it have that?
- 10:12Speaker 2: Is this section there? Is this calculation done? And just say, look,
- 10:16Speaker 2: this is the checklist I'd like you to have before
- 10:18Speaker 2: you deliver it to me. and it can actually review
- 10:20Speaker 2: and check on that checklist. You can even say, look,
- 10:22Speaker 2: tell me the checklist. And so if it hasn't hit
- 10:24Speaker 2: the checklist, sometimes it'll work harder until it actually gets
- 10:27Speaker 2: that checklist absolutely nailed. So give it that checklist, though.
- 10:31Speaker 2: Make sure it's clear rather than opinion-based. So is it good?
- 10:36Speaker 2: That's not a good checklist item. Is there a summary?
- 10:40Speaker 2: Is there an executive summary that's less than three paragraphs?
- 10:42Speaker 2: That is something that it can actually deal with. And
- 10:45Speaker 2: so you'll find that a checklist can give you a
- 10:47Speaker 2: better result without the AI going backwards and forwards and
- 10:51Speaker 2: you arguing with it so much. It can do its
- 10:53Speaker 2: own check effectively.
- 10:56Speaker 1: Neo, thank you so much for sharing five ways that
- 11:00Speaker 1: we can be changing how we prompt to get better
- 11:03Speaker 1: responses with the models that we are now working with
- 11:07Speaker 1: in late 2026. Thank you so much for listening to
- 11:12Speaker 1: today's episode of How I AI. If this show got
- 11:16Speaker 1: you thinking about how you could use AI better, that
- 11:19Speaker 1: is exactly what we help people do at Inventium AI.
- 11:23Speaker 1: We run AI training programs for organizations and also for individuals.
- 11:28Speaker 1: So if you want to upskill your entire organization or
- 11:32Speaker 1: maybe just yourself, you can find out more at inventium.ai. Hey,
- 11:37Speaker 1: and once more favor before you go, if you've got
- 11:39Speaker 1: a moment, like just five seconds, I'd love it if
- 11:42Speaker 1: you could leave us a review and hit follow wherever
- 11:45Speaker 1: you're listening. It really helps the show reach more people
- 11:48Speaker 1: and it helps us create even better episodes for you.
- 11:52Speaker 1: Thank you so much for listening and we'll see you
- 11:53Speaker 1: next time on How I AI. How I AI was
- 11:57Speaker 1: hosted by me, Amantha Imber and Neo Applin. A big
- 12:01Speaker 1: thank you to Martin Imber who does our sound editing
- 12:03Speaker 1: and Jem Rubio for production support. And thank you to
- 12:07Speaker 1: John Kilby who composed the theme music.