How to write a brief for AI: what it decides and what comes back to you
Before any steps, a brief for AI should draw the line between calls you leave to AI and calls that come back to you.
Key points
- Set goal and permissions before steps
- Write stop rules: contact, buys, secrets
- Set acceptance criteria before you ask
- Pass on revisit triggers and the reason
Contents
In the brief, set goal and permissions before steps
A&A perspective
When you hand AI a piece of work, write what the work is for and what AI may do before you spell out how to research it. Here, a brief means the short text you give AI for one job, covering four things: the goal, the permissions, the stop rules and the acceptance criteria.
A&A perspective
Ask AI to research something and it may come back with question after question, or go further than you asked. You may have settled the scope in your head, but if the brief does not draw that line, the checking and rework pile up.
From the sources
In a technical article for people who build AI that uses tools and works through tasks on its own, Anthropic names two common failure modes for instructions to AI.
From the sources
At one extreme, instructions hardcode complex logic and become brittle and hard to maintain. At the other, they give vague guidance with no concrete signals and falsely assume context the model does not share.
From the sources
The right instruction, the article says, is specific enough to guide the AI's behavior yet flexible enough to leave it room to judge. It also notes that keeping information minimal does not necessarily mean keeping it short.
A&A perspective
Automate & Augment (A&A), which writes this article, applies this to the single brief an owner writes. Leave the research steps to AI, and state the goal and the permissions plainly. The more steps you write in, the longer the brief gets, and the easier it is to leave out how far AI may go.
Hypothetical example
Say you are looking for a contractor to maintain your company website. Do not write the goal as "research contractors". Write "gather what I need to decide the order in which to interview candidates". The permission stops at reading public material and putting it into a table.
Hypothetical example
Once the brief says the final decision is the interview order, not an order to buy, AI has less room to go and ask for a quote halfway through.
Write stop rules, and calls that shift responsibility come back to you
A&A perspective
Stop rules in a brief make AI halt just before a call that is yours to make, so that call comes back to a person. A stop rule lists the moments where AI would otherwise make a decision that belongs to the owner.
A&A perspective
Draw the line at contacting outsiders, ordering or paying, and anything that needs non-public information. Permission to read material and permission to contact another company are separate permissions. If the brief says nothing, AI may send an inquiry form just to get closer to the goal.
From the sources
Anthropic's article on evaluating AI agents writes that two engineers reading the same spec can come away with different readings of how the AI should handle edge cases.
A&A perspective
A&A expects the same to happen between an owner and the AI they brief. Stop rules are where you write those edge cases down in advance.
Hypothetical example
In the contractor example, write two stop rules. One: "If a candidate does not publish its prices or scope, do not fill them in with a guess; return them as draft questions for the interview." Two: "Do not send inquiries or ask for quotes."
Hypothetical example
With these, AI will not rank a candidate on a guessed price when none is listed. It hands back a question to check in the interview instead.
A&A perspective
Add too many stop rules, though, and AI asks for confirmation every time, which defeats the point of asking. You do not need it to ask before each public page it reads. Stop only where something reaches outside the company, like contact or orders, and where an unsupported value would slip into the results, like filling a blank with a guess.
A&A perspective
Decide what AI returns when it stops, too. Have it return why it stopped and the question that, once a person answers it, lets the work go on. Then the owner only has to make the call.
| Situation | Who decides | What the brief says |
|---|---|---|
| How to research | Leave to AI | State the goal and permissions; do not script the steps |
| Reading public material into a table | Leave to AI | Do not make AI ask before each page |
| A candidate with no public prices or scope | Back to you | Do not guess; return it as a draft interview question |
| Sending inquiries or asking for quotes | Back to you | Forbid it in the stop rules |
| Ordering or paying | Back to you | Return why it stopped and the question to decide |
Set acceptance criteria for the results before you ask
A&A perspective
Write the criteria for accepting the results into the brief before you hand over the work. Acceptance criteria are the list of things to look at in the result before you can say the job is done.
From the sources
Anthropic's article on evaluating AI agents treats the record of the exchange with the AI separately from the actual state once the work is over. Its example: a flight-booking agent saying "Your flight has been booked" is one thing, and whether the reservation exists is another.
From the sources
For the tasks used to test how well AI performs, a good task is described as one where two domain experts would independently reach the same pass/fail verdict.
From the sources
It also says the quality of a research agent can only be judged relative to the task. One of the checks it lists is whether claims are supported by the sources.
A&A perspective
A&A reads this as: write acceptance criteria so that someone else checking the result would reach the same verdict.
Hypothetical example
In the contractor example there are four. Every candidate has the same comparison items, each item has a source URL, the unverified column has not been deleted, and AI has made no contact it was not allowed to make.
A&A perspective
If you accept the table because the ranking looks plausible, you can miss a table filled in with guesses. Looking at a proposed ranking and checking the evidence behind it are two separate steps.
Hypothetical example
When you ask for fixes, do not have AI redo everything; point only to the criteria it missed. For example: "Only candidate B has no source for its service area. Move it to the unverified column, or fill it from public material."
A&A perspective
With acceptance criteria written in advance, the check reads the same whether the owner or a contractor does it.
Pass on revisit triggers and the reason to the next brief
A&A perspective
Once you accept the results and decide something, pass the reason for the decision and the trigger for revisiting it to the next brief. Handing over only what the decision needs, rather than the whole previous exchange, keeps AI focused on the next job.
From the sources
Anthropic's technical article for people building AI says the amount of information AI can read at once should be treated as a finite resource. When Claude Code re-summarizes a long conversation, the article writes, it keeps architectural decisions and unresolved bugs and discards redundant output.
From the sources
It also writes that summarizing too aggressively can lose subtle details whose importance only becomes clear later.
A&A perspective
A&A picks what goes into the next brief the same way. In the contractor example, it is two sentences: "Candidates whose service area we could confirm from public material went first. If the maintenance scope does not fit, revisit the order."
A&A perspective
The whole previous exchange stays out because AI can read only so much at once. Hand over the detailed record of every candidate you compared, and the two sentences the next job needs get buried in it. Just as the summary keeps decisions and open issues, what you keep in the contractor example is the order you decided and the maintenance scope that could change it.
A&A perspective
Pass on only the reason, and the last conclusion gets treated like a fixed rule. Pass on the revisit trigger with it, and you avoid repeating the same debate while still allowing the review you need.
A&A perspective
If you are unsure what to pass on, check whether you can write the revisit trigger. A decision you cannot write a revisit trigger for does not yet have a clear reason.
A&A perspective
Keep the record of the previous exchange instead of deleting it, and reread it when a revisit trigger is hit. If the next brief is "write questions for the interviews with the top two", what you pass on is these two sentences and the candidate table.
Summary
A&A perspective
In the brief, write the goal and the permissions before the steps AI will take to research. Where contact with outsiders, orders or non-public information is involved, have AI hand the call back to a person through stop rules.
A&A perspective
Write acceptance criteria before you ask, and have AI keep the source and unverified columns in the results table. Pass the reason for your decision and the revisit trigger to the next brief together.
What to settle first in a brief for AI is not the steps but the line between calls you leave to AI and calls that come back to you. Next, decide where in the work you already hand to AI contact with outsiders, orders or non-public information should trigger a stop.
FAQ
Are there jobs where writing detailed steps works better?
Yes. For work that needs the same shape of result every time, like a monthly tally, or where the steps themselves are company rules, writing out the steps gives more consistent results. This way of briefing fits research and comparison, where the method can be left to AI.
With a lot of material, should I split AI into one that reads and one that compares?
Split only when there is a lot of material; with little, leave it to one AI. Anthropic's technical article for people building AI explains that an AI handed part of the work may explore a lot, but what it returns to the main AI is a summary. If you split, set in the brief which items come back, so exclusion criteria and unverified points are not lost in the summary.
Do I write company information and background into every brief?
No. Keep company information that stays the same somewhere else and hand it over separately. That setup is covered in another article, "Stop repeating yourself to AI: split company info into 3 parts". In the brief, write only what changes for that one job: the goal, the permissions, the stop rules and the acceptance criteria.
Sources & editorial note
Primary pages read for this article. Publication dates below belong to the sources; access dates record our research.
- Effective context engineering for AI agents
Anthropic · 2025-09-29
Accessed 2026-10-24 - Demystifying evals for AI agents
Anthropic · 2026-01-09
Accessed 2026-10-24