Transcription
If you want to make the most of Notion AI, you need to get one thing right: your master prompt. It's the difference between having an assistant that just gets you and one that requires constant handholding. Without it, you will constantly have to micromanage AI, and you will probably get frustrated and give up.
So, in this video, I want to walk you through:
A. Why do you need a master prompt in the first place?
B. The two archetypes of master prompts.
C. And then, last but not least, how to set up your own master prompt in less than 5 minutes.
Plus, I also share one thing that you need to get right if you're serious about Notion AI that has nothing to do with Notion itself. Ready? Let's go.
So, first, why do we even need a master prompt? Well, it comes down to the first rule of AI: garbage in, garbage out. You need to write good prompts and give AI context if you expect it to do the job well. But we're inherently lazy. Who really has the time and wants to tell AI every single time, "Well, here's where tasks go, and this is what we should do if X happens"?
That's where a master prompt comes in. It's essentially preloaded context, a set of instructions that AI references every single time before it executes a task. That means instead of repeating yourself, you can get away with much shorter prompts and still get really good outputs.
Plus, if you work with a team, it's really essential to make sure everyone uses AI the right way. You know how hard it is already, right? To make sure you always prompt things the right way. And through a master prompt, right, you can make sure that generally the AI behaves the same way across the company for every team member.
But what kind of master prompt should you set up? There are essentially two types of master prompt approaches. There's the MMP, the multi-mode prompt, and then there's the CLP, the context layer prompt.
In a multi-mode prompt, you essentially try to anticipate most of the tasks that AI might have to do for you and then give it very specific instructions on how to identify what kind of task you're currently asking it to perform and then what instructions it should use when doing so.
A context layer prompt, or CLP, on the other hand, is all about establishing the general high-level guidelines and then using specific instructions whenever you need them and pulling them in on demand.
That might sound a bit abstract. So, here are two examples of how this could play out.
Here we have a multi-mode prompt example. And the very defining aspect of this type of general instruction is that at the very top, you explain to the system, "Hey, there are multiple modes to choose from. Um, whenever a user makes a request, this is how you choose what mode you should operate in, right?" And it's actually fairly similar to how GPT-5 works, right? You might have noticed the change from four to five. In five, you have a lot less models to choose from, and by default, you type something in, and then there's a first lightweight model who makes a decision: "Okay, is this something for GPT thinking? Is this something for fast response?" And multi-mode prompts try to emulate the same effect.
So, in this case, right, just for this example flow, we would have our workspace manager, a content writer, maybe a meeting processor, and a research assistant. And then all of these would have additional information inside, right, here. Of course, this is just an empty page, but you get the gist. And then you have some selection guidelines, right? When should a specific element be picked? And then you might have some global rules as well, right, that apply independent.
But the core, right, is really this idea: "Hey, when a request comes in, decide first which way to route it."
A context layer prompt, on the other hand, skips this routing and instead focuses on giving the AI the general context that it needs to operate in our system. And you can think of it as if, you know, a new person joins the company, and you explain them, "Hey, here's how we work." So, in enough creating this, you also probably get better at onboarding people, just as a side note.
Basically, we have some very, in this example, very quick run-through of like, "Okay, what is it that we actually do? We are a Notion consultancy, right, that work with companies and teams and help them in our 8-week Notion transformation sprint to create the foundations they need to use Notion ideally. Here's our core workspace architecture. Right? These are the main databases that we use, and so on and so on." And then typically explain high-level workflows, why they are always relevant, like how do we do project management and how do we do knowledge management, and then some key behaviors and guidelines with the AI. So, overall, this prompt is a lot more lightweight, right, and more hands-off when it comes to specific detailed processes, but it gives a quick brief on, "This is how work gets done."
So, which one is right for you? I would say a multi-mode prompt is only really applicable if you work on your own and you have a very sort of defined set of spheres or areas that you work in. In pretty much every other situation, I highly recommend that you go with the CLP, the context layer prompt.
And there are a few reasons for that. First, with a multi-mode prompt, it will take you a lot more time to set things up because you need to think through all the different, you know, types of work that it might have to perform for you, right? So, it will take you a lot longer to create a thing, which in turn makes it less likely that you'll ever get to it in the first place.
Second, it requires a lot of foresight, and you need to really know what kind of work you actually have to do. For most of us, what's on our plate will change quite regularly. So, anticipating every major scenario will be nearly impossible.
Three, it just doesn't scale to a team environment at all. It's complicated enough to try to anticipate for yourself what you might have to do, but it's impossible to do that for everyone in the organization. And even if you could, it would water down the instructions for the individual use case so much because there are so many things in there, right? And only a small part is applicable to the individual person. Plus, right now, everyone needs to set up their own instructions, which makes it very, very hard, right, to maintain this across the board.
Number four is a small but meaningful one, and that is that a multi-mode prompt increases response time. If you have a lot of instructions and you ask AI to always sort of review a large chunk of it and then make a decision of what's best for that situation, that means it has to go through these calculations every single time. Now, that's fine for complex tasks. In fact, for complex tasks, you want AI to think as long as possible. But for all these small little workflows throughout the day, like "add a task here," "create a quick note there," right? You want it to be just snappy and respond quickly.
And number five is my main argument for a context layer prompt. And that is that you don't want to let AI choose the tool for the job at hand. This is one of the areas where you, as the human, have so much value to add. You know what needs to be done. You know, right, what the situation calls for. Your judgment is the essential part, and then you have AI do the execution. In a multi-mode system, right, you sort of like dilute this a little bit because you tell AI, "Hey, you know, here's a situation I need help with. Now, please go and figure out how to approach it." And that again just adds more variety to the potential outcomes. So, I like to be in control here, right? And pick exactly how AI should solve a given problem for me.
Now, in a context layer prompt, we still have specific instructions for specific situations, but—and that's the crucial part—they live outside of our master prompt. Our master prompt is designed to explain to AI, in the shortest and most succinct way, "This is our workspace. This is the work we do. These are the typical everyday tasks that I will ask you to perform." And then, when it comes to specific procedures, they live in separate pages. They are documentation both for humans who might have to do the task and for the AI that is asked to do it, and you can reference and inject them whenever you need them.
Again, I see that this might be a bit, you know, abstract in theory, but let's look at an example of how this works in a day-to-day workflow.
The first example here is about everyday operations, right? Task management. You need to create them after a meeting. And on the left, you see the prompt that you would need to write if you want to get this done well, reliably afterwards, without a context layer prompt. You need to explain it: "Hey, go through these notes, right? Which database should it create these in? How should it name the task? Right? What is your, what is your naming convention around that? What about statuses? What is the default? What is something that it needs to keep in mind? Due dates, do you always add them? Right? Is there a default? Is the team allowed to have tasks without a due date or not? How does it find out what project it should be related to? Right? Is that something it can infer from information in the meeting? Um, where should additional context go, right? Should it always add context to the page body or not? Um, plus, right, like generally, what other elements do you want to have it linked with?" And again, this is something that you probably do several times a day. So, you can't, right, if you have to write this every single time you want, right, that's that's just the nature of things, you will skip. And that way, that means AI will start making assumptions, and your outputs will be less efficient.
But with a context prompt, right, all these things are part of your general instruction of how work gets done at your company. So, all you need to do is say, "Well, please create tasks from this meeting." If you want to, you can add "in accordance to your general instruction," but you could probably even leave that out, right? Because it knows what it needs to do to create these tasks.
Now, let's look at a more specific workflow. My team and I have started to go through our meetings, particular ones with clients, and extract the key metaphors and frameworks that we use because there are few things, right, that we keep coming back to, that we keep referencing, and we want to build out our library so that we can easier share that with each other. That's of course a lot of work, and without AI, pretty unrealistic to do. With AI, it's still a night and day difference whether you need to instruct it or whether you have it built into the system.
So, without any instructions, right, we need to tell it: "This is the purpose, right? I want you to go through it, extract metaphors, frameworks, etc. Here's what we mean, right? What is sort of like a metaphor? What is a framework? What is useful? How do you, how do I want it to extract these? Where should it add them? Right? And what should it do there?" Whereas if we extract once how this process works, then this is the actual prompt that we use in the company to do this, right? We explain it exactly that it should do this in this database, it has the criteria, it has a naming framework, it has everything, right, that it should and shouldn't do, and it lives outside of our main master prompt because again, this is maybe required once or twice a week, so I don't want to overload my prompt in general with that or have to update it. All we need to do is reference, "Hey, you know, follow procedure 87A," and we'll go ahead and do that.
See the difference? Everyday operations like task management live in your context layer prompt. Where to create a task? What are the minimum properties? And what should you ask me in case I don't provide you with that information? Specific procedures like the framework extraction live in their own individual pages and can be injected whenever the situation calls for it, which is your decision, not the AI's. This keeps your context layer prompt fast and lean, and responses snappy, and still gives you the opportunity to have really specialized processes.
So, now let's go to the next big chapter. How do you actually create a context layer prompt? For this part, we are going to follow my standard outline for a context layer prompt. It's the one that we used to create the master prompts for pretty much all our clients, and it's a great starting point. Now, of course, you can adjust this to your specific situation, right? So, if there's something that you don't need or want to add to it, feel free. It's more about the spirit, right? This idea of giving AI a general overview of what you do and how you do it. Also, if you want, you can get this as a free template, right? And I have a link down below to join my newsletter. And if you're on the list, you will get an email with a ton of different templates, and among them, this context layer sort of follow-along, fill-it-out page.
To start with, assuming you use Notion for work, the first thing you need to do is tell it about that work, right? What's your company? What do you do? Who do you serve? Right? The quick rundown of what it is that you do. In here, in this example, we also have, in case you have a small team, right? You could list that out with role and responsibilities so that for specific tasks, it can tag the correct person. Or if you have a larger company, add your team directory, your employee directory, list out who does what so that if it's necessary, it knows where to find it.
Then, second, workspace architecture. Very important. Explain what are your core databases, what are they used for, and how do they interact, right? How do they connect? This of course requires you, right, to have a proper setup. So, if you don't, this is your chance now, right, to go through it and actually create this series of global databases for your company. And then again, just like you would explain a new person who joins your team, "Well, this is how we do it," right? You explain to AI, "Right, this is where that goes, and this is what we use this database for."
With your company and database structure explained, you want to go through the general way that you work, which in most companies will mean explaining your project management and knowledge management approach. Right? This might be as simple as, "We use tasks and projects. This is where tasks go. This is where projects go. If we create a task, these are the properties that typically should be filled out." Right? Or if you have a more complex system, for example, right, let's say you have like a three, four, or even five-level hierarchy between maybe initiatives, goals, epics, tickets, something, if in, you know, very engineering-driven companies, then explain in that, tell it where something should go, and where in your system the different moving pieces of everyday work fit in.
This is also a chance to reinforce the team culture that you want to build. For example, let's say you want to make sure that no tasks fly around in the system without a due date. Well, tell the AI that if it's obvious what due date should be set, right, it should infer that from the instructions, or if not, it should maybe always set the due date for next week, or maybe it should always explicitly ask the user, right, and remind them that, "Hey, no tasks without a due date, what do you want to set?" That's where AI really is super helpful because you can bake the general guardrails, right, that you want to have for your system into the interaction with the user.
For knowledge management, it's similar. One very important point here, right? Our general recommendation for pretty much every company that we work with is that you want to have a general docs database that acts as a catch-all for all the pages created in the company because you want to avoid that people start working in their private area, right? Just create random pages in the sidebar and then maybe later forget to move them to the right place. That's why we always say, "If it doesn't go into a specific dedicated database, right? If it's not a task, not a project, something like that, it should always go into this one catch-all database." And again, with AI, you can kind of enforce it, and you can tell AI, "Whenever you create a new page, unless you know it belongs in a different place or unless the user specifically asks you to create it elsewhere, always create it in that place." That is already in general a great rule because Notion AI has a tendency to create new pages as top-level pages in your private area, which is definitely not where you want to have things for proper organization.
Every company needs project and knowledge management. And once you're done with that, you might also want to outline some key cross-functional components of your company. For example, if you have a very specific OKR system that requires everyone to report on a weekly basis, maybe it makes its way also into the context layer prompt. Or if you have very specific product management, right? For example, you track your features and sub-features in Notion. And you want to make sure that when tickets in the task management, right, or documents in the knowledge management are created, that they link always to that right product. You might want to spend a few minutes here explaining how that works. But of course, we don't want to get too detailed here, right? It's not about our Notion choosing and routing. So, if you feel like, you know, you need to write pages and pages here, then it's much better to reserve that right for these external pages, and here maybe give a brief overview of typical workflows and where to find more information.
And then it's pretty much back to a normal prompting where you outline key behavior and guidance, right? General tone, style, operating rules, things like, you know, making sure that when it, um, when a user asks it to create a database, make sure to first check the back end, right, in our global databases, whether that maybe exists because we don't want to end up with two task databases. Things like that. And last but not least, right, maybe an agent identity, explaining it what it's generally expected it to do. These two are less important, right, in the bigger scheme of the context layer, because Notion AI will default to, you know, sort of a good personality here, but if you want to steer it in a specific direction, this is your chance to do so.
Once you're done creating this, it's time to roll it out to your company, and that is at the moment still a little bit tricky. So, just want to briefly touch on it. Basically, you want to create this general prompt in the public area in your workspace, ideally probably in your knowledge management system, right? In your docs database, then shared with the rest of the team.
Now, once shared, everyone will have to go in and add this to their own workspace. Uh, sorry, to their own instructions because we currently can't set a company-wide master prompt. Doing so is fairly straightforward, though. Very simply, right? You ask everyone to copy the whole content of the page, then add a new page in their private area. This is one of the few cases, right, where it's okay to have a page in the private area. Paste the instructions in there, and then go to Notion AI and set this, right? You can set this by clicking here on "Personalize," and then you have the option to "Edit your instructions" and set a certain page as that. That's the easiest default, right? And make sure that you have, you know, the same instructions rolled out to everyone, and then everyone can go from there and start adjusting it to their own specific needs.
That of course has one drawback, and that is in case you update your instructions, users will have outdated versions, and you need to ask them to update it again. So, to avoid that, the easiest way, right, is to add a sync block in your page, right? Then copy the whole context of your page, right? Let's take all of this here, and then move it all into this one sync block. And now, what you can ask users is you can ask them to simply copy the content of that sync block, add it to their private page, and then whenever you change something here to the company-wide system, right, it will automatically update in their places as well.
Important here, only to keep in mind to make sure that you then reduce the permissions on this page so that everyone else in the company can only view this. Because if you have full edit permissions on this, that means anyone could change the master prompt for everyone else in the company, and that will guaranteed to happen in larger organizations, right? So, you're going to avoid that one person can mess it up for everyone else. And by setting the permissions on the source to "can view only," that means that only also on every instance of the sync block, people will only be able to view it, right? Notion permissions always follow the origin of where something comes from. If you have issues with that, I have a full tutorial on Notion permissions linked down below in the description.
Now, at this point, let's quickly talk about this one thing that you definitely should do that has nothing to do with Notion. To make all your AI usage much better, you need to invest in an AI dictation tool. AI dictation tools are simple. You just press a key, start talking, and it transcribes everything that you say with a far better accuracy than all the built-in dictation tools. It's the single biggest unlock because it allows you to add so much more context and speed up things like creating your context layer prompt massively. You can simply copy the template, right, that I have for this master prompt, add it to your workspace, pull up Notion AI, hold down the dictation key, and tell Notion AI, "Hey, it's time for us to set up this master prompt. Here you see the structure. Now, let me tell you a little bit about how we work." And then you just go point by point, right? And you quickly explain it. "These are your main databases. This is how you like to add tasks," and so on. It's a matter of 5 minutes or less rather than having to type everything down manually.
Which one of these dictation tools you pick matters a lot less than that you get one in the first place. Now, my current favorite is Monologue, and I have an affiliate link for that down below in the description in case you want to support the channel. But really, Monologue is amazing. I had Whisper before, also really great. And as long as it does what it says on the box, right, powerful, fast dictation, then you're good to go.
Now, one thing that you might have noticed when looking at the blueprint, right, for the context layer prompt is that you need to have a good and solid Notion foundation. If you don't have a proper workspace architecture with clearly defined databases and rules, then all your Notion AI efforts will go nowhere. If you need help with building out that foundation, then my team and I would love to take you through our signature process, the 8-week Notion transformation sprint, which is specifically designed to make sure that everything is AI-ready and works perfectly well for the humans involved. If that sounds interesting, just send me an email or check out the link down below in the description.
Now, with your context layer prompt set up, it's time to slowly but steadily expand on your list of specialized procedures that you can call upon whenever needed. And when it comes to setting those up, I like to use a framework called ACDC. I recently spoke about the AC/DC method at my keynote at Make with Notion in Munich, and I will record a longer, in-depth YouTube video about that soon. So, if you want to know exactly, right, how to expand the first step, then make sure to subscribe to the channel so that you don't miss that video. But that's it for now. Everything you need to create the one prompt to rule them all.
Now, I'd love to hear from you. How have you set up Notion AI? Do you use a master prompt already? And if so, are you team multi-mode prompt or team context layer? Plus, if you've been playing around with Notion AI for a while, you might have run into some issues already. Things like when you ask it to process a large amount of data, it keeps timing out. Well, I have six amazing Notion AI prompting tips in this video next, which among other will walk you through the best solution to get Notion AI right to process a lot of things autonomously without you requiring to go in all the time. Just click here, and I will see you in a few seconds.