An EA in one of the bigger admin communities ran a small experiment last week.
She took a task shaped exactly like her own daily work, finding public information about a business from the general details she had, and handed it to an AI.
The AI came back and said it was unable to answer reliably from the information provided, and it needed more to work with.
Her reaction, in her own words:
She meant it as a complaint. She is right to. Underneath the complaint sits the most useful thing anyone said about AI all week, and almost nobody caught it.
The machine did not fail. It told her exactly what was missing.
You have never once had that option. And the skill you built because you never had it is the same skill that makes these tools work.
Subscribe to The Dispatch and the rest of this issue unlocks right here, along with the full archive. One email every Tuesday, about 8 minutes, built for admin professionals. No spam, unsubscribe anytime.
Already subscribed on this device? This unlocks automatically.
Your AI output is mediocre because of the brief. Not the prompt wording.
Anthropic’s engineering team describes the shift plainly. Working with these models has moved away from finding the right words for a prompt, toward deciding what configuration of information the model should be looking at when it answers. Not phrasing. Context.
Which is a sentence you already knew, about people, ten years ago.
Think about what happens when a director drops a request on you with three words and no context. You do not fail. You go get the missing pieces first. Who it is really for. What the deadline is. What happened last time. What the exec will object to. Then you execute.
The model has no way to do that step. It will not go and find the missing pieces. On a good day it tells you they are missing. On a normal day it fills the gap with a plausible guess and hands you something confident and generic.
That is where the generic comes from. There is a Senior EA in the same community right now, ten years of C-suite experience, who has paid for resume services, used AI to strengthen her resume, and is still not getting interviews. Buried in her own post is the question answering itself: are my bullets too polished, too generic, too AI-sounding? Yes. The tool was handed a resume and asked to improve it, and nobody gave it the thing making her different from four hundred other Senior EAs. It had no way to go and get that. It is not in the document.
This is not a prompt template. It is the same intake you already run on humans, written down so you run it on the machine too.
Who is doing this work and what standard are they held to. “You are an executive assistant to a CFO at a 400-person company. Everything you write goes out under her name.” This is not decoration. It sets the register, the caution level, and the vocabulary.
Not “help me with the offsite.” “Draft the pre-read for a two-day leadership offsite.” One verb, one output.
This is the whole ballgame and the part everyone skips. Who the audience is. What was decided last time. What the exec hates. The internal history. The politics. Paste the real material, the previous email, the last agenda, the actual notes. Not a summary of them. Them.
Length. Format. Tone. Deadline. What is off the table. “Under 300 words. No bullet points, she hates them. Do not mention the vendor situation.”
One line, and it changes everything. Put it at the end of every brief:
If you are missing something you need, ask me before you write.
That instruction hands the model the option you never get. Use it. It converts a confident wrong answer into one good question.
Take the next request you get that is genuinely underspecified. The three-word ask. The “can you just handle this.”
Before you touch it, write the brief. All five parts. Paste in the real prior material.
Two things happen. The output gets dramatically better, which is the small win. And you end up holding a written statement of everything the task actually required, which is the large one, because that document is the case for what your job involves in a form you hand to somebody.
Then hit reply and tell me what the model asked you for. The question it came back with is usually the thing nobody told you either. I read every reply.
The brief makes the model useful on the parts of your job that are retrievable. It does not make the model good at your job.
Judgment about which of two directors will be offended is not retrievable. Knowing your exec is short on Mondays is not retrievable. The reason the vendor relationship is delicate is not retrievable. None of it sits in any document, which is precisely why it is the part nobody automates.
One more practical thing. These tools handle information at the start and the end of a long brief better than information buried in the middle. If something is critical, lead with it or close with it.
The model gets to say I need more information. You get one shot and no clarifying question.
So write the brief anyway. Write it for the machine, and notice you have just written down the invisible half of your own job. The part living in your head, the part nobody documents, the part no one hands off. That is not prompting. That is the thing you were already good at, finally on paper.
Master your tools. Own your work.
— Kristy
Digital Zena subsidiary of Fusion AI Consulting LLC
The AI Prompt Library for Executive Assistants, the Meeting Prep and Follow-Up Template, the New Tool Adoption Checklist and the rest of the Digital Zen resources are free and open. No form, no cost.
Open the library →