Write a brief an agent can finish

1. Name the artifact and the decision it supports

Begin with what you want to have when the agent stops. Name the artifact, its audience, and the decision or action it should support. ‘Research our competitors’ can lead to a large report with little practical use. ‘Create a two-page comparison for our product meeting so we can choose which onboarding pattern to prototype’ gives the work a purpose and a useful size.

Specify the destination when it matters. A reply in chat, a local document, and an update to a shared workspace are different outcomes. Include the format if another tool will consume the result. For human readers, describe the information hierarchy before dictating cosmetic details. A concise recommendation, supporting evidence, and unresolved questions usually matter more than a particular heading style.

2. Supply the context that changes the answer

Include the facts that a competent newcomer would need. Explain the current process, the specific problem, relevant constraints, and which source is authoritative. If two documents disagree, name the one that governs the task or ask the agent to surface the disagreement. A pile of attachments without a reading order leaves the agent to guess which details carry weight.

Prefer a short set of relevant materials to an unexplained archive. Label each item by purpose: current policy, example output, raw data, or background. Add dates when freshness matters. If you provide an example, say which qualities to copy. Otherwise the agent might imitate a sample's accidental length, stale claims, or formatting mistakes. Keep private details out when a small representative sample will answer the question.

3. Separate decisions the agent can make from fixed constraints

Give the agent room to choose a useful method while making the important boundaries explicit. For a memo, it can group findings and decide which comparisons deserve space. It cannot invent evidence, change the source proposals, or send the recommendation to the team unless you authorize those actions. For a coding task, it can investigate the relevant files while preserving unrelated work.

State the boundaries that are plausible in this specific task. A page of unrelated prohibitions makes the actual requirements harder to find. Include the allowed work area, any external actions that need review, and the limits on time or scope. If you want exploration before implementation, say where that phase ends and what decision you will make next. A request to propose an approach should not silently become permission to execute it.

1. Name what may change

   Identify the document, folder, branch, draft, or test workspace where the agent can act.

2. Name what must stay intact

   Call out existing content, interfaces, data, or decisions that the work must preserve.

3. Name the consequential handoff

   Specify whether publication, sending, purchasing, deployment, or another external action belongs in the task or requires a later decision.

4. Make completion observable

Write acceptance checks that you can actually perform. ‘Make it excellent’ describes an aspiration. ‘Include all three proposals, cite the supplied evidence for each estimated benefit, and mark missing information as unknown’ gives the agent a target and gives you a review method. Use a small number of checks that capture the core job rather than an exhaustive wish list.

For behavior changes, describe a setup, an action, and the expected result. For research, require the evidence behind the claims that determine the recommendation. For a transformation, specify what must be preserved and how omissions should be reported. Also define incomplete success. If one source is inaccessible, the agent should produce the verified portion and identify the missing piece rather than present a complete-looking answer.

5. Tell the agent how to handle uncertainty

Some missing details are cheap to resolve with a stated assumption. Others change the direction of the task. Give the agent a rule for the difference. It can choose a reasonable memo structure and say what it chose. It should ask before selecting between conflicting business objectives or acting on an ambiguous account. This keeps minor decisions moving while protecting the choices that belong to you.

Set a sensible stopping rule. For a bounded research task, ask for a first pass within a time budget and a list of remaining questions. For a blocked tool connection, ask the agent to preserve completed work and name the access it lacks. Avoid instructions that reward endless activity. Your brief should make a candid partial result more useful than a long chain of repeated attempts.

6. Improve the brief after one real run

Read the result against your original checks. When it misses the target, diagnose the cause before adding more words. Was the source missing, the instruction ambiguous, the permission unavailable, or the agent unable to do the work? A stronger adjective cannot repair a missing input. Correct the narrow cause and keep the rest of the brief stable so that you can tell whether the change helped.

Save the useful parts as a reusable template: outcome, inputs, boundaries, acceptance checks, and uncertainty rule. Leave task-specific facts as clearly named fields. Remove instructions that no longer matter. A good brief becomes shorter as you understand the job better. Keep one accepted example alongside the template, and update that example when your standards or underlying process change.