Brief agents like a manager.
Most people prompt. Managers brief. A prompt is a wish typed in a hurry; a briefing is what you’d give a bright new hire on their first day — enough intent to steer by, enough context to decide with, enough constraints to stay safe, and a clear picture of what done looks like.
Most people prompt · Managers brief
A wish typed in a hurry. The agent fills every gap you left with a guess.
Intent to steer by, context to decide with, constraints to stay safe, and a clear picture of done.
The gap between them isn't ten percent. It's the difference between an intern guessing and a chief of staff executing.
The idea
A prompt is a wish typed in a hurry. A briefing is what you'd give a bright new hire on their first day: enough intent to steer by, enough context to decide with, enough constraints to stay safe, and a clear picture of what “done” looks like. The difference in output quality is not ten percent — it's the difference between an intern guessing and a chief of staff executing.
Match the brief to the stakes. A throwaway task gets a throwaway prompt; a recurring or high-stakes one earns a brief. It's the same discipline as 2. STAFFING & BUDGET in the template below — you match the model to the stakes, so match the effort you spend briefing to them too.
Write the brief once per recurring task, refine it every time work comes back wrong, and keep your briefs in a folder — that folder is your management system. When output disappoints, resist fixing the output. Fix the brief. A rebrief beats a redo, every time, because the brief compounds and the redo doesn't.
The template
Copy this. Fill every line. Short is fine; blank is not — a blank line is a decision you're forcing the agent to make for you.
Copy this — fill every lineAGENT BRIEF — [task name] Date: [ ] 1. ASSIGNMENT One sentence: produce [what], for [whom], so that [outcome]. 2. STAFFING & BUDGET Agent/model tier: [frontier / mid / fast-cheap] — because [stakes] Budget: [$ or token cap] · Deadline: [ ] (You wouldn't pay a principal engineer to rename files. Match seniority to stakes. Know what this task SHOULD cost.) 3. INTENT — the why What larger goal does this serve? Who consumes the result, and what will they do with it? What does "great" look like? (Link an example if one exists.) 4. CONTEXT — what the new hire needs to know Background docs / links / prior decisions: What's been tried already, and why it didn't stick: House style, conventions, vocabulary to use or avoid: 5. CONSTRAINTS — the guardrails Do not touch: [systems, files, topics, claims] Must include / must never include: Scope line: this task is NOT [the adjacent thing it will drift toward] 6. DEFINITION OF DONE Acceptance criteria (checkable, not vibes): [ ] [ ] [ ] Format & length of deliverable: 7. REVIEW PLAN — trust and verify, by stakes Checkpoint(s): [e.g., outline before draft; schema before migration] Escalate to me immediately if: [ambiguity, missing access, conflict] Review depth: [skim / standard / line-by-line] — because [stakes]
A worked example
The same brief, filled in for a real recurring task. Notice how little each line needs to say — the value is that no line is left blank.
The same brief, filled inAGENT BRIEF — Weekly competitor digest Date: Mon 1. ASSIGNMENT Produce a one-page digest of competitor moves this week, for the GTM team, so Monday standup starts informed instead of surprised. 2. STAFFING & BUDGET Mid-tier model — synthesis task, low blast radius. Cap: $2/run. Deadline: Sunday 6pm, recurring. 3. INTENT Serves: quarterly goal of sharper competitive positioning. Consumers: sales + product leads; they'll raise anything material. Great looks like: last week's digest (linked), minus the fluff. 4. CONTEXT Competitor list + why each matters: [link] We ignore funding rumors; we care about pricing & feature moves. 5. CONSTRAINTS Do not speculate on unannounced products. No claims without a source link. NOT a strategy memo — observations only. 6. DEFINITION OF DONE [x] Every item has a dated source link [x] One "so what" line per item, one page total [x] Explicit "nothing material" section if a competitor was quiet 7. REVIEW PLAN No checkpoint (low stakes). Escalate if a competitor announces pricing changes. Review depth: skim.
The five briefing sins
Every disappointing agent run traces back to one of these. Each is also a way to mismanage a person — which is the point.
- 1The blank why
An agent without intent optimizes the literal words of the assignment. So does a confused employee.
- 2Context hoarding
You know things the agent needs. Every one you withhold, you’ll pay for in revision.
- 3Vibes-based done
“Make it good” is not acceptance criteria. If you can’t check it, you can’t delegate it.
- 4Overstaffing by default
The most powerful model on every task isn’t rigor; it’s a payroll you don’t understand.
- 5Fixing output instead of briefs
The redo dies with the task. The rebrief improves every future task.
Where your briefs live
A brief isn't a one-time write. The review outcome feeds back into it — so it's a living document that gets sharper every cycle. Kept together in a folder, your briefs become the operating layer between you and everything you delegate.
That folder is the same idea as the Work OS field guide — a private management system as a folder of markdown. The briefs are how you hand work out; the Work OS is how you keep track of what comes back.
Keeping each brief current — and carrying it between you and the agents — happens to be the part Ari does on its own.
More field guides
Copy-paste playbooks you can run todayA private, file-based chief of staff you run in Claude Code — the TRAM loop as a folder of markdown.
A prompt pack that turns Claude into an org-health analyst — feed it your data, get the 21-point scorecard and spider chart.
The hand-it-to-Claude-Code build spec for a personal AI work OS — vision, a legible data model, a full prompt pack, and a phased build plan.
A fifteen-minute weekly review on the FIT loop — Feel, Inventory, Target. The 1:1 you owe the first report on your own org chart.
The brief, carried for you.
Keep your briefs in a folder and you have a management system. Ari is what carries the brief between you and the agents — keeping intent, context, and definition of done current as work moves.