The outcome-first brief: give an AI agent the result, not the steps

Two ways to brief an AI agent

Most people writing their first agent brief write a task list: research X, summarise Y, draft Z, attach A. It looks organised. It feels thorough. And it often produces worse results than a single well-stated outcome — because it pre-empts the sequencing, limits the agent's discretion on approach, and bakes in assumptions that may not hold once the work is actually running.

This pattern is drawn from Ethan Mollick's July 2026 guide to which AI to use for non-experts, published in One Useful Thing. Mollick draws a clear line between back-and-forth chatbot use — fine for quick single-step tasks, almost any model will do — and agentic delegation, where you hand off hours of work in a single turn. The key move in agentic delegation, he argues, is writing the desired end state rather than the steps, and then stepping back while the agent sequences its own work.

The live example Mollick shows: connected to Gmail, the instruction given was "Help me prep for the MBA seminar I'm giving on Monday the 21st, including building some presentation and demos as inspiration. Answer any outstanding messages on the topic." The agent returned completed teaching materials and drafted email replies in roughly ten minutes. No task list. One outcome sentence.

The task-list brief tells the agent how to move. The outcome-first brief tells the agent where to arrive. The distinction is not cosmetic: a task list transfers your sequencing assumptions to the agent, which then executes the list rather than the goal. An outcome brief hands the sequencing to the agent, which may find a better path — or reveal that your assumed path was missing a step you hadn't named.

The prompt

Paste this at the start of an agentic session. Replace the brackets. Use Claude or ChatGPT with tool access enabled (email, calendar, files, or relevant business apps). The outcome-first frame works with any agent that can execute tasks rather than just generate text.

Why each piece earns its place

"Act as a capable colleague, not a chatbot." This line changes the default interaction mode before anything else happens. Chatbot mode produces clarifying questions before acting. Colleague mode produces decisions — with reasoning — and then action. The shift alone saves several back-and-forth exchanges on straightforward preparation tasks.

Naming the tools explicitly. Agentic sessions require tool access to produce real output. "Connect to Gmail" is not metaphorical: the agent needs that integration enabled before the session starts. Naming the tools in the brief also forces you to decide what surface you are actually delegating, which surfaces scope problems early rather than midway through.

One outcome sentence. This is the hardest discipline in the brief. Most people write two or three desired outcomes, which the agent may prioritise in an order you wouldn't have chosen. One sentence forces you to decide what done actually looks like. If you genuinely need multiple outcomes, order them explicitly: "First X, then Y if time remains."

The "name the choice, then proceed" rule. Without this rule, agents halt on any unresolved ambiguity and ask a clarifying question. The halt breaks flow and returns the decision to you — which is exactly what delegation is supposed to avoid. This rule gives the agent permission to decide while keeping a visible trace of every decision made, so you can review the reasoning afterwards.

The irreversible-action gate. Outcome-first delegation hands the agent more discretion than a task list. The gate draws the line: it may decide freely up to the point of consequence. Sending, publishing, deleting, and financial actions pause for your confirmation. Everything else, the agent handles.

The three-part handoff. Agents do not naturally surface what they chose not to do. "Completed" is obvious. "Skipped" is not — and skipped items often include the thing you cared about most but stated least clearly. Requesting it explicitly closes that gap, and the pattern of what gets skipped tells you what to add to the outcome sentence next time.

Failure modes

The vague outcome. "Help me prepare for Monday" is not an outcome sentence. The agent has no way to know what "prepared" means. Be specific: "I need a rough presentation outline, drafted replies to the three most recent emails about the session, and a list of open logistics questions."

The invisible context problem. The agent works from what it can access. Internal context — the relationship history with a client, an unwritten constraint about a team member's availability, the political background on a project — is invisible unless you put it in the brief. Front-load any context that changes what "done" means.

Accepting the first pass. Outcome-first delegation is not outcome-free review. The agent's first pass reveals its assumptions about your priorities. Read the handoff, check the "skipped" section, and use the review step before treating the session as complete. The gap between what you expected and what came back is the useful signal, not a failure to be ignored.

No tool access. The brief describes work the agent physically cannot do without connected integrations. Test access on a trivial task first — confirm the agent can read your inbox, open your files, or see your calendar — before running a consequential brief that depends on those surfaces.