This is the kind of advice about using AI I’ve found myself repeatedly sharing with my smart friends who are not software engineers and still figuring out how to get the most out of AI.
They are using AI, but haven’t figured out how to wield this new sword effectively yet. Typically, they’ll be using ChatGPT instead of Google Search, and maybe using some AI tools like NotebookLM.
They might say things like “I asked it to make a presentation and some of it was surprisingly good but a lot was terrible and I couldn’t use it. Not sure if that’s me doing something wrong or AI just isn’t good in my industry yet?”
Bad news: By now, it probably is you, not the AI.
Good news: Anyone can learn a few techniques to help AI deliver better for you.
Here are my starter tips that’ll get you 80% of the way (we’re still figuring out the rest):
My Mental Model for AI. Think of it like a really smart, driven intern, that shows up on day 1 and has no sense of what you like or how the workplace works. Unlike a smart intern however, it won’t naturally learn over its first few weeks. Every day it shows up, is its first day. So you have to get good at rapid onboarding.
It can’t read your mind. Frontier models are amazingly capable now (and getting better). The gap is they just don’t have enough context about you, your situation, and what you’re asking. If you ask it “make me a cake”, it’ll go off and do everything and there’ll be a cake, but it probably isn’t going to be what you had in mind.
Include rich examples. It will respond well to examples, in addition to clear instruction. “I want a vanilla and blueberry sponge cake for 4 people. No icing, use butter cream. No fancy decoration, just plain white with a couple of blueberries on top. Here are the recipes of the last three cakes we made which we loved bc they were light and not too sweet: (paste recipes)”
Analogy: Think of when your boss asked you to do a thing you haven’t done before. And you get the general idea, but it’s a bit vague. Talking more doesn’t help too much, you’re itching to just see an example of what was done before to better understand it.
Make (and refine) a plan first. For a task of any reasonable complexity, always ask it to make a plan first. Maybe your AI has a built-in plan mode for exactly this, but if not, you can just say “I want a cake. Research and make a plan of what cake you will make and how you’ll make it. Ask me questions you need, and lets review together before you start”. Then review and tweak the plan so it is closer to what you want. Then let it start your ask to “make me a cake”.
Analogy: Think of when you rush out and do something and discover along the way, there’s actually lots more details to figure out. Or when you plan a party, there’s actually way more details to consider once you flesh out the plan that you first thought.
Use other AIs to cross-check. Before you lock in your plan, you give the proposed plan to another AI and ask it to review the plan and what questions/specs it thinks are missing to do a good job. Playing AI against AI like this helps a lot. Gemini, Grok, Claude, ChatGPT, Qwen, etc all have their own strengths and weaknesses, and most have a free tier you can use for this.
Analogy: When experts give you advice in a domain you don’t really understand, it’s pretty common to get a second opinion. Or asking a room full of people to align on a plan will surface different concerns/issues from different perspectives. Unlike a room full of humans, you can ask the AIs to review, critique and improve the same plan as much as you want.
Break big work into sub tasks. Then have it complete those one by one and review/approve or give feedback to the AI after each one. You can just ask it “let me review and approve when you’ve made the base and again when you made the icing, before we put it all together”. Usually making the plan will naturally lead to a breakdown of subtasks, but if not, ask for it to get broken down into useful milestones or review points.
Analogy: you’d break down a complex task into a few key steps for a new hire. Importantly though, you’re not just solving for the AI who will likely make small mistakes along the way that will compound and might make the final result seem bonkers. You’re also solving for your ability to review - it’s easier to inspect each part and confirm it’s what you expect or course correct than look at the final product and have this feeling that the overall thing is wrong but struggle to itemize all the issues at that point.
Practice, be curious, and ask for things you feel it can’t do (it might surprise you). You really just need to spend time practicing these techniques, and you’ll develop your own taste and intuition for it. When the AI does something wrong or seemingly silly, don’t write it off and think AI is stupid, just look around at what context they had and what info they were missing to have got it right. For example, sometimes at the end of a project, I’ll say: “now that’s done, give me a full detailed prompt that I could give to an AI to recreate this in one-shot?” And by looking at what it produces, I get a sense for what details mattered. And now I can see what I really needed to say instead of “make me a cake”.