For most of the early period with AI tools, the default pattern was to treat every interaction like a search query: compress the request into one or two sentences, hit send, then edit whatever came back. The results were fine. Sometimes good. Rarely exactly right on the first pass.
What gradually becomes clear is that the gap between “fine” and “useful” is usually not the model’s fault. It is the way the work was framed.
The pattern among people who get consistently good results is not that they write cleverer prompts. It is that they stop prompting, in the search-query sense, and start briefing instead. A brief is what you would give a smart new colleague before asking them to take on a real piece of work: the actual outcome you need, the context they would lack, the constraints that matter, and what done looks like. A prompt is a line of text. A brief is compressed intent.
The six fields that make a brief work
A useful way to think of a complete brief is as covering six fields. None of them are complicated on their own — the discipline is in not skipping the ones that feel obvious.
Goal is the first and most often skipped. The common mistake is to write the activity (“help with a summary”), not the outcome (“a two-paragraph summary that a non-technical CFO can use to decide whether to fund this”). The model cannot meet an unstated goal — it will infer one, and it will not always infer yours.
Context is what a smart colleague joining late would need: who the audience is, what has already happened, why this is happening now, what the reader already believes or is worried about. Not a data dump — the facts that would change the answer.
Sources adds a hierarchy that often goes unused: not just “here’s a document,” but “this is authoritative, this is background, this is outdated, don’t use this.” That shift turns the model from a text generator into something operating against evidence.
Constraints are the boundaries that prevent technically correct but practically useless output. “Don’t invent numbers.” “Don’t change the voice.” “Don’t draft a formal proposal yet — give me a one-paragraph summary I can take to the first conversation.” These are not about making the model timid. They preserve the real-world conditions the output has to live in.
Quality bar is where taste lives, and it is the field most consistently under-specified. What does “good” mean for this piece of work? Who is the toughest reader, and what would satisfy them? The model genuinely cannot meet a bar you have not shared.
Definition of done tells the model what form to return, when to stop, and whether there is a checkpoint before continuing. This is best understood as a form of mercy — it is the difference between a converging process and an endless one.
Name the edges, not just the centre
There is a posture point that sits behind all six fields. Think of it as a flashlight: name both the centre of the beam (your intent) and the edges (what is out of scope). Most of us skip the edges. In my experience, telling a model what to leave out is often more useful than telling it what to put in. Scope exclusion turns out to be one of the highest-leverage things you can put in a brief.
When the overhead is worth it
Not every task needs all six fields. A quick lookup, a casual question — none of this warrants a full brief. The calibration principle is that overhead should match the stakes. Quick exploration stays loose. Recurring workflows, anything public-facing, hiring decisions, customer communications, financial analysis — these reward the investment in a complete brief, and the return on that investment compounds when the workflow repeats.
There are also helper prompts that operationalize the six fields: a brief-builder that walks you through them conversationally, an auditor that takes a vague ask you’ve already written and rewrites it field by field, and a definition-of-done generator for when the task is clear but you can’t describe what should come back. Of the three, the auditor is the one worth trying first.
Where this points
The shift from prompting to briefing reflects something real about how the practice is maturing. The early phase was figuring out how to talk to the model. The current phase is figuring out how to delegate to it — which requires the same discipline we would apply when delegating to a person. The brief is the instrument of that delegation.
The models keep getting better at filling gaps. But the brief is still our job.
Text summarized and optimized using Anthropic’s models and reviewed by a human.