Create Flawless Skills/Prompts with the “DRAW” Method
Today I’m here with one of the hottest topics: “How to Guide Your Agent Correctly”. This topic is very special to me because it is actually related to the bet I made years ago about being a generalist. Just to be clear, I want to start by defining the word "generalist".
Being a Generalist and Integrated Intelligence
A person who is a “generalist” knows about most topics in terms of "high-level overview, the why, usage of the concept, alternatives, tradeoffs, and actionables related to the topic”.
Being a generalist is like being able to paint a clear picture of a concept, knowing neighboring concepts and how all of these actually relate together, how a topic can be useful in real-life, and knowing its examples… but most importantly, with a tendency of not going too deep into the implementations.
Apparently, when you’re a generalist and focus your learning on that particular approach, you feel smarter and smarter every day, but also dumber and dumber in the sense that without AI you may sometimes feel like an imposter. The reason why you’re feeling smarter and smarter is because this particular way of learning benefits your “integrated intelligence”, thank God they have a word for it.
Integrated intelligence: Refers to the practice or capacity of synthesizing multiple distinct forms of intelligence, information, or processing modes into a unified, cohesive whole.
So you may not be able to write a function in Python that perfectly implements DFS (exaggerating a bit, or am I?), but at least you can come up with the most innovative ideas on where to use DFS. Because you know why it’s useful, how it works at a high level, why you should use it, and where you should use it.
To Write Better Prompts, Be a Generalist
Let’s get to the point: the reason I actually mentioned being a “generalist” is because you actually have to be a generalist in order to guide your agent better.
But why? Because if you can actually draw a flawless picture, articulate what you want better than everybody else, and have the creativity to connect topics and capture patterns like no one else, with a sense of vision, constraints & limits, and a specific focus on DoD (Definition of Done), then you’ll probably do a better job than anyone else at that specific task.
The reason why I tell everyone to become a generalist with phrases like “just know a topic at a very high level, learn where to apply it, and be aware of example use cases, and you’re done” is because “learning how a specific implementation is done” is becoming less and less meaningful every single second (for most of the topics).
Therefore, change your way of learning: just focus on learning different aspects, broader and at higher levels. I’m aware that what I’m saying is very controversial, and surely there are counterexamples. If you want to become a research scientist in a specific topic or actually enjoy deep-diving into that topic, be my guest.
But if you actually do not enjoy learning that topic (but you have to) or you won’t become a researcher in it… don’t force yourself to learn every single detail. First, it will burn you out, and second, it’s unnecessary and a waste of time because nowadays, there are few tasks that an incredibly smart model cannot do (such as Claude Opus 5.5 or GPT-6 Astra).
How Being a Generalist Actually Helps Write Perfect Prompts
Now, according to recent research from OpenAI, Anthropic, and many arXiv papers (referenced at the end), you have to let the agent decide how to work a specific task out and not hinder it. It’s actually more understandable in this case:
“Imagine you have the most intelligent person in the world who can implement anything, would you actually tell them how to do that task? Or would you just give specific details and let them decide, since they are more intelligent than you?”
It’s very simple, right? All you have to do now is paint the right picture, as flawless and complete as possible, and let the agent connect the dots and fill in the blanks for you. Since it’s more intelligent than you, already has nearly all the necessary information, or knows how to retrieve that information better than you, your job is just to oversee its work (as HITL) and draw the lines of the frame.
Being a generalist helps you specifically with that, because it helps you draw flawless and complete pictures. The reason? Very simple: since prompting is now about giving an agent the bigger picture, if you (as a person) have tuned your learning toward understanding the bigger picture in any kind of topic, you’ll be able to draw bigger, more complete, and more flawless pictures than anyone else. It’s basically a “practice makes perfect” kind of thing.
The DRAW Method
So DRAW is not a new innovative method that I came up with but nobody named this so I did. And since it’s also about drawing a better, bigger, flawless picture for agents in a specific tasks, it does the job.
Above are the only 4 aspects that you should include (omit the others) in your SKILL.md, AGENT.md, CLAUDE.md… files. Think of them as buckets rather than a strict list. Examples of good output belong under DoD, because they show what done looks like. Exact commands for fragile steps belong under Rules, because they're boundaries the agent shouldn't cross. With that clear, let's dig into DRAW:
Definition of Done (DoD)
Starting from a philosophical perspective, DoD is arguably one of the most important concepts in life. As humans, we need a our own definition of meaning, so even our lives have a built-in Definition of Done (or at least we try to find it). But DoD isn’t just for living creatures or research teams; it’s also essential for any task at hand. So, when you define a clear DoD for a task and give that to an agent, it now has a way to evaluate itself, allowing it to ask: "Am I done?" It’s like navigating a crowded street: you know the exact spot you need to land on, but you have to zigzag through the crowd, jump over a few obstacles, and maybe even take side steps… When things are complex, things inevitably go wrong, but as long as the agent has that clear end picture in mind, it can self-correct.
Rules (Constraints + Never-Dos)
Having the constraints said out loud, telling it what it should never do, limits the agent within a closed frame. You give it boundaries, but being within boundaries does not mean it’s no longer free. It just means that whenever it tries to find its way, it won’t cross these boundaries. An example constraint would be “cost”: knowing the “cost” should be under a few dollars means it can now limit its options from 300 to 3. I ask you, isn’t it now freer than ever? Sometimes having fewer choices actually helps you achieve the goal (just like in life; imagine you were dealt all the cards, wouldn’t it be impossible to define a meaningful life?).
Aim
It’s about what we are physically trying to build or achieve in this specific task. People often confuse Aim with DoD, but a simple clarification solves all the question marks. The “Aim” is the direction, and the DoD is the destination. It’s like telling a person: “You should follow this direction to land at that exact spot.” For example, you should “draft a product email” (Aim), and it should include a catchy subject line, 3 key feature bullets, and a single call-to-action link (DoD). That’s when you know you’re done!
Why
We have told our agent: “You have to get a pen because we’re drawing, you cannot use a red pen, and this is the only place on this blank paper you can draw on”. Now we’re telling our agent why we’re drawing in the first place. Giving the why builds an even bigger picture: how does this specific action contribute and connect to a bigger cause? I think it’s very self-explanatory and easy to understand why we’re giving our agents “the why”. Practically, how this turns out is that our agent can even come up with a completely innovative alternative. If we say, “The reason we’re drawing is because we want to have fun,” now it has the awareness to say, “If you want to have fun, play a video game,” and create an alternative, a version, or an opportunity that you’ve never thought of.
So What’s the Action to Take Here? (Two Actionables)
Most blogs end up having no actionables at the end, but that is not my style of writing (maybe it’s because I’ve been a Stoic over the past few years).
Now, I have two actionables for you (which I apply myself):
-
First, create a daily reminder for yourself (just for a month or so) that says “DO NOT FORGET THE DRAW METHOD”, so that when the time comes and you are in front of your PC writing a prompt, you actually remember to write in this format.
-
Second, revise your existing skillset that you use with your agents and all the .md files that you have (SKILL.md, AGENT.md, CLAUDE.md…). Skills are just template prompts that are discoverable and implicitly or explicitly callable to be injected into the context window.
Alright, I know it was a bit of a long read. I hope that this blog benefits you in some kind of practical way, because it did for me (writing it was a lot of fun). Take care until the next read.