Session Slides

We build agents with the best intentions, but they disappoint us. We assume they understand context they don't have, or we give them instructions so vague they could mean anything. The good news is that agent performance is not a mystery. It's entirely predictable. If you want your agent to do something consistently, you have to write instructions the way you would write them for a new hire who can never ask you a follow-up question. That sounds harder than it is. Today we're going to show you exactly how.

This distinction matters more than anything else you'll learn today. A prompt is a one-time instruction. You ask ChatGPT something, get an answer, move on. A brief is different. A brief is the permanent operating manual for an agent that will run dozens or hundreds of times. If you write a prompt when you should be writing a brief, your agent will either fail silently or do something nobody wants. We're going to teach you how to write a brief, not a prompt.

Think of these four components as the skeleton of every agent brief that actually works. Context is about giving your agent the background it needs to make good decisions. It's not just the task, it's the why. Task is the action itself. Get this wrong and everything fails. Constraints are the guardrails. If you don't tell your agent what to never do, it will try everything. Output format is where most briefs fail completely. Engineers and managers skip this part because it feels tedious. But if you don't show your agent exactly what success looks like, it will guess. And it will guess wrong.

Context is not filler. Context is the difference between an agent that makes one mistake and an agent that makes the same mistake every time. If you tell your agent it is a financial analyst, it behaves differently than if you tell it it's a writer. If you tell it the audience is a regulatory body, it behaves differently than if the audience is a startup board. The more specific your context, the fewer bad assumptions your agent will make. Don't be vague about who your agent is or who it's talking to.

Task is simple on paper but hard in practice. You have to be specific about what the input looks like and what you want back. If your input is a document, is it always a PDF? Is it always English? Does it have a maximum length? On the output side, if you ask for analysis, do you want bullet points or paragraphs? Do you want sources cited? These details matter. On constraints, most teams either skip this section or write vague things like don't be biased. Instead, identify the specific things your agent could do wrong, and forbid them by name. If your agent processes financial documents, explicitly tell it not to make up numbers. If it talks to customers, explicitly tell it not to promise things the company doesn't offer.

Output format is not optional. It is not a suggestion. It is the only way your agent can know if it has done the job right. If you tell your agent to analyze something but don't show it what the analysis should look like, it will invent a format every single time. You'll get inconsistency, wasted cycles, and agents that look broken when they are actually just confused. The strongest briefs include an example of a real output that came from the agent itself. Show your agent what good looks like.

This is the hardest part of writing a good brief, and it's where experience matters. You have to imagine all the weird inputs, edge cases, and corner scenarios your agent might encounter. Then you have to write instructions for each one. If you do not tell your agent what to do when something is ambiguous, it will guess. And when an agent guesses, it often picks the wrong option. The best way to identify edge cases is to run your agent on real data and see where it fails. Then write instructions for those failures.

Testing your brief is not optional. You have to run it before it goes live. Give your agent the inputs it will actually see, not just the easy ones. If it fails on any input, rewrite the brief and test again. The brief is not done until your agent can handle the real world, not just the ideal case. Once it passes all four tests, you are ready to hand it to your team with confidence.