Recently, AI engineers started circling the same idea from different directions. OpenAI’s /goal. Anthropic’s Dynamic Workflows and Outcomes. Less “do these ten things,” more “here is what success looks like.” Define done, let the system work, and verify when it arrives.

Good creative directors have known this for thirty years. It’s called a brief.

The brief is not a preamble. It’s the finish line, dressed up as a starting gun. A photographer who shows up to an event without a shot list comes back with twelve hundred photos of everything. With a shot list, she comes back with a hundred photos of what you actually needed. The camera is identical. What changed is whether anyone defined done before the session started.

What the engineering tools made visible is a failure mode that most knowledge workers experience daily but rarely name. The failure is happening in the thirty seconds before you open the chat, when you’re carrying a vague sense of what you need and haven’t made it specific. The AI surfaces that vagueness in four minutes. That’s its actual gift, in a slightly uncomfortable direction.

OpenAI’s /goal works by giving an agent a finish line, a verification method, and constraints. The agent loops around that finish line until it reaches it or tells you it’s stuck. A recent Codex guide from Every put the practitioner version of this well: describe what you want to end up with, not how to get there. Their test is useful beyond coding: if you’d type the same sentence into three consecutive prompts, it belongs in the goal, not the chat.

I am not suggesting every team should use /goal. It is built for engineering tasks with measurable benchmarks, and most knowledge work does not behave that cleanly. The lesson for knowledge work is not “use this exact feature.” It is that serious AI work now starts with success conditions, constraints, and verification.

What used to be bundled into a good brief now has to be made explicit. What are we trying to produce? How will we know whether it worked? What has to be understood before anything gets made? AI does not eliminate those questions. It punishes you faster for skipping them.

So I’ve started treating the brief (aka my prompts) as three separate jobs: define done, define good, and define the problem before anything gets made.

The first is defining done. Before I open a chat, I write the finish line from the perspective of the person who will use the output. “A research synthesis that gives the head of strategy three reasons to green-light the vendor change, written for someone presenting it to the CFO in forty-eight hours” is a finish line. “Something about the vendor options” is a wish. The difference is five minutes and the willingness to think about your audience before you think about your content, which is, embarrassingly, what every good brief has always required.

I started applying the Every team’s test and immediately noticed how much of my prompting is re-briefing mid-session because I hadn’t written the brief first. The AI didn’t drift. I started without a destination.

Defining done tells you where you’re going. Defining good tells you whether you arrived. Anthropic is experimenting with something they call Outcomes, which separates the agent that generates from the agent that evaluates. The evaluator gets only the finished output and a rubric. No access to the reasoning that produced the work, no attachment to the choices made. It grades against criteria alone. I started applying the same separation to my own process before I’d heard the name for it.

Knowledge workers have said the same thing for years: you cannot evaluate your own work from inside the process that made it. You know what you meant. That knowledge gets in the way of seeing what’s on the page.

So I write the evaluation criteria before generating anything. Before I draft a piece, I write down what makes it succeed: the reader understands one new thing, that thing connects to a decision they’re facing. It changes what I ask for, and it changes how I evaluate what comes back. When AI is generating, the rubric gives you language for revision that isn’t reactive. “This doesn’t meet criterion three, which says the strategic implication needs to appear in the first three paragraphs” produces different results than “this feels like it’s missing something.” One is a direction, while the other is a mood.

Writing three to five observable criteria before anything exists is enough to change the quality of everything that follows. The brief becomes the scorecard, and the scorecard was always what the brief was for.

The practice I resist most is running a planning session before any generation begins, which probably means it’s the most important. Claude Code has a feature called Planning Mode that enforces exactly this: before the AI touches anything, it reads, analyzes, and surfaces assumptions. It cannot write or change anything. Fast generation without that review produces what engineers call technical debt, which is a polite name for a beautiful facade on a shaky foundation you now own forever.

Knowledge work has technical debt, too. We call it “this campaign doesn’t hang together” or “this strategy is beautifully written and completely wrong.” It shows up after the generation, when you’re trying to fit pieces together and discovering they were never going to fit.

So I run the planning conversation first. The AI’s only job is to analyze the problem. What are the tensions in this brief? What assumptions am I making? What does the audience need to believe already for this argument to land? What would make this fail? I sit with those questions before we touch anything. Sometimes it changes the brief entirely. Either way, the generation that follows is faster and more aligned than anything I produced by opening a chat and hoping for direction.

Engineers needed agents to run for hours without losing the plot. Knowledge workers needed teams to make work that held together across strategy, design, and execution. Different vocabulary, same underlying problem: the brief is the work. Everything after it is execution.

Execution now takes four minutes. Which means the brief matters more, not less. The taste embedded in it, the judgment about what success looks like, and the discipline to define the problem before producing the work. Those are what’s left.

Before you open your next session, write one sentence describing what a successful output looks like from the perspective of the person who will use it. Then write three criteria for knowing whether it worked. See what changes. Because the next frontier of AI work is not better prompting, but better briefing.

— Lauren Eve Cantor

thanks for reading!

if someone sent this to you or you haven’t done so yet, please sign up so you never miss an issue.

I’ve also started publishing more frequently on LinkedIn, and you can follow me here

if you’d like to chat further about opportunities or interest in AI, or this newsletter, please feel free to reply.

banner images created with Midjourney.