How to Write a Task Brief an Agent Can Finish
A task brief that an agent can finish fits on one screen. That is the first test, and it eliminates most of what gets handed over as instructions. Briefs fail in a predictable way: they describe a mood, a goal, a hope, and leave the actual work undefined. A capable agent will attempt the task anyway, guess at the missing pieces, and return something that has to be redone from scratch. The hours lost are not the agent's fault. The brief was never finishable.
The word "agent" here means a worker, a contractor, a junior colleague, or a software system that takes an instruction and produces a result without asking follow-up questions. The brief has to survive that silence. Everything the worker needs must be on the page before work begins.
The difference between a wish and a brief
"Improve the onboarding email" is a wish. It names a feeling about a document and no destination. An agent reading it has to invent the definition of improved, the audience, the length, the tone, and the deadline. Each invention is a guess, and guesses compound.
A finishable brief reads more like this: rewrite the onboarding email so a new customer understands the first three setup steps, keep it under 200 words, match the plain tone of the account settings page, deliver it by Thursday. That version answers five questions before anyone opens a text editor. What changes, why, how it should read, how long it runs, and when it lands. The difference is not polish. It is whether the work has an edge to stop against.
Most briefs collapse because the writer assumes shared context that does not exist. The agent was not in the meeting where the new pricing was discussed. The agent has not seen the last six versions of the document. The brief has to carry that freight, or the result will drift.
Name the deliverable in physical terms
Agents finish best when the output has a shape they can picture. A brief should specify the artifact: a spreadsheet with these seven columns, a two-page memo with a recommendation on page one, a list of twelve vendor names with contact details, a slide deck of eight slides ending on a cost estimate. Physical terms beat abstract ones every time.
Compare "research the market" with "produce a one-page table of the five largest competitors, their published prices, and one sentence on each pricing model." The second version tells the agent what to count, what to compare, and when to stop. Research with no stated output is a hole that swallows days.
The count matters as much as the format. "A few examples" invites three or ten, and the agent will pick whichever number feels generous. "Three examples" is a decision already made. Every number in a brief removes a decision from the worker and puts it where it belongs, with the person who understands the task.
Constraints do more work than instructions
A brief with open constraints is a brief that never ends. The agent can keep improving the draft forever, because nothing says when it is done. Constraints are the walls that make finishing possible.
Useful constraints are specific and checkable. A word limit, a budget ceiling, a deadline, a list of sources that are off limits, a file format, a required section order. Each one is a line the agent can confirm it has not crossed. Vague constraints ("high quality," "professional tone") do not constrain anything, because no one can verify them at the end.
There is a temptation to over-constrain in the other direction, to write a brief so tight that the agent has no room to solve problems. That produces brittle work. The right balance is firm on outcomes and loose on method. Say what the result must be, and let the agent decide how to get there. A brief that dictates every keystroke is a brief that should have been done in-house.
State what done looks like before work starts
The strongest single sentence in any brief is the one that describes the finished state. "Done means a single document that a new hire could follow on their first morning without asking a question." That sentence tells the agent what standard to aim for, and it tells the reviewer what to check. Without it, every review becomes a negotiation about taste.
Definition of done works best when it is concrete enough to be tested. "Done means the spreadsheet totals reconcile to the source file" is testable. "Done means the report is insightful" is not. The test does not need to be elaborate. It needs to be something the agent can run against its own output before handing it back.
This is also where the brief should list the acceptance criteria as a short checklist. Three to five items is plenty. Each item should be something a stranger could verify in under a minute. If an item takes an hour to check, it is not a criterion, it is a second project.
Give the agent the reference material it cannot guess
Most task failures trace back to missing inputs, not missing skill. The agent did not know the brand voice because the style guide was never attached. The agent used the wrong customer list because the right one lived in a folder no one mentioned. A brief should carry its own supplies.
Attach the minimum set of references and say what each one is for. "The attached style guide governs tone. The attached spreadsheet is the only source of customer names. The previous version of this memo is attached for structure only, not for content." That last distinction matters, because agents tend to over-copy examples. Labeling what to borrow and what to ignore keeps the output from turning into a clone of the sample.
When a reference does not exist, say so explicitly. "There is no prior version. Build the structure from scratch." That sentence prevents the agent from hunting for a template that was never there, or worse, inventing one and presenting it as the standard.
Write the brief in the order the work happens
Briefs read badly when they are organized by topic instead of by sequence. The agent has to reconstruct the order of operations from prose that jumps around. A better structure follows the work: what the task is, why it exists, what inputs are provided, what the output must contain, what constraints apply, and how it will be judged. That order answers questions in the order they arise.
The "why" deserves a sentence of its own, early. An agent that understands the purpose can make better judgment calls on the small unstated decisions that always come up. A brief that explains the goal behind a pricing comparison will get a more sensible comparison than one that just says "compare prices." Purpose is not decoration. It is the tie-breaker the agent uses when the brief runs out.
Build in a stopping point for questions
No brief is complete. The honest ones admit it. A good brief includes a short section on what the agent should do when it hits an ambiguity: stop and ask, or make a reasonable assumption and flag it in the deliverable. Picking one of those two rules in advance prevents the two worst outcomes, a stalled task and a silent wrong guess.
For most work, the better rule is to proceed with a flagged assumption. Telling an agent to stop at every gap produces paralysis. Telling it to guess silently produces surprises. Telling it to guess and list the guesses on a final page gives the reviewer something to check without blocking the work. The brief should say where that list goes and how long it can be.
The two-page ceiling and why it holds
Briefs longer than two pages stop being read carefully, by people and by agents alike. Length is a signal. A brief that runs to eight pages suggests the writer has not decided what matters, and the agent inherits that indecision. The discipline of fitting onto two pages forces the writer to separate the essential from the merely relevant.
Two pages is enough for a title, a purpose sentence, a deliverables list, an inputs list, a constraints list, a definition of done, and a rule for ambiguity. That is the whole apparatus. Anything beyond it is usually background that belongs in an attachment, or hedging that belongs in the writer's own head.
The ceiling also creates a useful editing habit. When a brief spills past two pages, the overflow is almost always specification creep: the writer trying to control method instead of outcome. Cutting back to the essentials improves the brief and shortens the review cycle at the same time.
A brief that survives the agent's absence
The real test of a brief is what happens when the writer is unreachable. The agent has the document, a deadline, and no way to ask a clarifying question. Does the work still land? If the answer is uncertain, the brief is not finished. It is a draft with good intentions.
A finishable brief is a small, self-contained machine. It carries its own purpose, its own inputs, its own limits, and its own test for success. The writer's job is to build that machine and then get out of the way. The agent's job is to run it. When both sides hold up their end, the handoff is quiet, the review is short, and the work comes back done.
