The Art of the Brief
A great agent outcome is mostly decided before the agent starts. The brief is your highest-leverage move — here's the anatomy of a great one.
🎯 By the end you'll understand
The five parts of a strong agent brief — goal, context, constraints, success criteria, and optional resources/format — what specific failure each part prevents, why concrete examples beat adjectives, how to right-size a brief, and why iterating on the brief is normal, not a sign you got it wrong.
KEY INSIGHTThe one idea to take away
You don't steer an agent mid-flight. You aim it before launch. The brief is the aim.
In Lesson 3.1 you saw that talking to an agent isn't chatting — it's delegating. The brief is how you delegate well. Almost every disappointing agent run traces back to a thin brief: the agent did exactly what was asked, but what was asked was vague, so "exactly" landed nowhere useful.
Here's the leverage. A few extra minutes spent sharpening the brief routinely saves you an hour of re-runs, corrections, and "no, not like that." The agent will faithfully amplify whatever you hand it — clarity or confusion. So the brief is the single highest-return thing you can improve about how you work with agents.
An agent can't read your mind, and it won't push back the way a good colleague does. Everything it needs to succeed has to be in the brief or reachable from it. Treat the brief as the contract — write it like the outcome depends on it, because it does.
DEFINITIONWhat a brief actually is
A brief is the instruction you give an agent at the start of a task: not just what you want, but enough context, boundaries, and a clear definition of done that the agent can work toward your real intent without you hovering.
It's the difference between handing a contractor "make the kitchen nicer" and handing them a one-page spec: new white shaker cabinets, keep the existing layout, budget £4k, done when the doors are hung and the old ones are hauled away. Same person, same skill — wildly different odds of getting what you actually wanted.
A brief is a structured instruction that gives an agent a concrete goal, the context it needs, the constraints it must respect, and the success criteria that define "done and good." It's a contract written for someone who is capable but literal, and who won't ask clarifying questions unless you build that in.
Note what a brief is not. It's not a prompt-engineering incantation, and it's not a wall of polite filler. It's a clear specification. The skill is plain, and you already have most of it — you just have to apply it on purpose.
FRAMEWORKThe anatomy of a strong brief
A strong brief has up to five parts. The first four are the load-bearing ones; the fifth is a frequent bonus. A handy way to remember them: Goal, Context, Constraints, Criteria — the four C-ish pillars (goal is the odd one out, but it's first for a reason).
1 · Goal — the outcome, stated concretely
What does "done" produce? State it so it's checkable. Not "help me with the report," but "a 2-page summary of Q2 sales by region, as a Markdown file." A goal you can't verify is a goal the agent can't aim at.
2 · Context — what the agent needs to know
Background, where things live, and the relevant specifics of your situation. "The data is in sales_q2.csv; regions are EMEA/AMER/APAC; this is for the board, so keep it plain." Context is the difference between a generic answer and one that fits your world.
3 · Constraints — the boundaries
What not to do, plus limits on budget, time, style, scope, and tools. "Don't touch the raw data file. No external API calls. Under 400 words. Use British spelling." Constraints are how you stop a capable agent from over-reaching.
4 · Success criteria — the definition of done
How both of you know it's finished and good. "Done when: the file exists, every region has a figure, totals add up, and it reads in under 2 minutes." This is the agent's checklist — and yours.
5 · Resources & format (optional)
Tools it may use, files it may read, and the exact shape of the output. "You may use the calculator tool; output as a table followed by 3 bullet takeaways." Often folds into context, but calling it out removes guesswork.
Goal + Context + Constraints + Success criteria (+ Resources/Format). If your instruction is missing one of the first four, you've probably found the reason your last run went sideways.
CORE CONCEPTEach part prevents a specific failure
The framework isn't bureaucracy for its own sake. Each element exists because leaving it out causes a predictable, named failure. Once you can name the failure, you'll spot the missing part instantly.
| Brief part | What it gives the agent | What's missing it causes |
|---|---|---|
| 🎯 Goal | a target to aim at | Drift — it produces something plausible but not what you needed |
| 🗺️ Context | your specific reality | Generic / wrong output — answers a stranger's version of the task |
| 🚧 Constraints | edges it must stay inside | Over-reach — touches things it shouldn't, balloons in scope, wrong style |
| ✅ Success criteria | a finish line | Never-done — stops too early, or loops forever unsure if it's enough |
| 🧰 Resources/format | means & shape | Guesswork — invents tools/format you then have to redo |
The two most expensive omissions are goal and success criteria, because they bracket the work. A vague goal means the agent starts in the wrong direction; missing success criteria means it never knows when to stop. Get those two crisp and you've fixed most bad runs before they happen.
When a run disappoints, don't just re-prompt louder. Ask: which of the five parts was thin? Drifted off-target → sharpen the goal. Touched the wrong thing → add a constraint. Stopped half-done → write explicit success criteria. The fix is almost always a missing brief element, not a smarter model.
SEE ITBrief Builder — watch a weak brief become strong
Start with a one-line brief and a low strength meter. Toggle each component on and watch the brief grow, the strength climb, and the predicted outcome flip from likely to drift to likely to nail it. Scoring is fixed per component — no magic, just the framework.
PRACTICALShow, don't just tell
Here's the upgrade that punches above its weight: examples beat adjectives. "Make it professional" means nothing — your "professional" and the agent's "professional" are different planets. A single example collapses that ambiguity instantly.
Compare. "Write a friendly, concise product update." versus pasting one past update you loved and saying "match this tone and length." The second leaves almost nothing to interpretation. Adjectives describe a vibe; examples define a target.
Adjective (weak): "Use a clean, modern table."
Example (strong): "Format like this — | Region | Revenue | Δ vs Q1 | — one row per region, currency with thousands separators, sorted by revenue descending."
This applies everywhere: tone (show a sample), structure (show the skeleton), code style (point at a file to mimic), output shape (paste the format). When you catch yourself reaching for an adjective, ask: could I show an example instead? Usually you can, and it's worth the ten seconds.
CORE CONCEPTRight-sizing: enough, not a novel
A common over-correction after learning this framework is to write briefs the length of a legal contract. Don't. Right-sizing means including exactly what's needed to remove ambiguity — and nothing more.
More words are not more clarity. A bloated brief buries the load-bearing instructions under filler, and an agent reading 600 words of preamble may weight a throwaway aside as heavily as your real goal. (You'll see in Lesson 3.3 that the agent's attention and context window are finite resources — padding spends them.)
For each sentence in your brief, ask: "If I deleted this, could the agent still get it right?" If yes, cut it. If deleting it would let the agent reasonably guess wrong, keep it. Aim for the shortest brief that leaves no room for the wrong guess.
The right size scales with the task. "Rename these files to kebab-case" needs one line. "Refactor the auth module" needs goal, context, constraints, and criteria spelled out. Match the brief's weight to the stakes and ambiguity of the job — not to a fixed template you fill out every time.
COMPARISONBefore & after: one brief, rewritten
Let's make it concrete. Here's a weak one-liner, then the same request rebuilt with the five parts. Same task, same agent — but only one of these reliably produces what you wanted.
"Summarize our Q2 sales and make it look good."
No checkable goal, no context (what data? who's it for?), no constraints, no definition of done. The agent will produce something — and you'll spend the next hour saying "not like that."
Notice it's longer — but not much longer, and every line earns its place. There's nothing left for the agent to guess about. That's the whole game: spend a few minutes here, save an hour of correction later.
You will rarely nail the brief on the first try, and that's fine. When a run misses, the brief just told you which part was thin — tighten it and re-run. A brief is a living draft, not a one-shot exam. Good operators expect two or three rounds and budget for them. (More on this rhythm in Lesson 3.5.)
WEAKNESSCommon mistakes to avoid
Length isn't clarity. A padded brief buries the load-bearing instructions and spends the agent's limited attention on filler. Right-size: the shortest brief that leaves no room for the wrong guess. Cut any sentence the agent could succeed without.
It won't be. Without a definition of done, the agent either stops too early or loops forever, unsure if it's enough. Always state how both of you will know it's finished and good. It's the cheapest part to write and the most expensive to omit.
"Clean," "professional," "concise" describe a feeling, not a target. The agent's interpretation won't match yours. Show an example instead of stacking adjectives — paste a sample, a format, or a file to mimic. One example beats a paragraph of describing.
KEY INSIGHTWhere this connects in your course
- Talking ≠ chatbot (3.1) — the brief is the mechanics of delegating, the shift you met last lesson.
- Context engineering (3.3) — the brief's "context" part scales up into a whole discipline: what to load, what to leave out, and why attention is finite.
- Iterating & supervising (3.5) — briefs are living drafts; this is where you learn the re-run rhythm and how to catch a drifting agent early.
- Planning & goals (2.5) — a clear goal is what lets the agent decompose the task into a sensible plan; a fuzzy goal yields a fuzzy plan.
The brief is where Stage 1's "agents are context, not magic" becomes a daily habit. Master it and most of the frustration people have with agents simply disappears.
GLOSSARYKey terms introduced
- Brief
- The structured instruction given to an agent at the start of a task: goal, context, constraints, and success criteria.
- Goal
- The concrete, checkable outcome the task should produce — stated so you can verify it's met.
- Context
- The background and specifics the agent needs about your situation: where things live, who it's for, relevant facts.
- Constraints
- Boundaries the agent must respect — what not to do, plus limits on scope, budget, time, style, and tools.
- Success criteria (definition of done)
- The explicit checklist that tells both you and the agent when the work is finished and good enough.
- Right-sizing
- Including exactly enough in a brief to remove ambiguity — and no more. The shortest brief that prevents the wrong guess.
- Drift
- When an agent produces something plausible but off-target, usually because the goal was vague.
- Over-reach
- When an agent does more than intended — wrong scope, wrong files, wrong style — usually because constraints were missing.
ACTIVE RECALLCheck your understanding
Answer in your own words first, then reveal.
1What are the five parts of a strong brief, and which two matter most?›
Goal, context, constraints, success criteria, and optionally resources/format. The two that matter most are goal (a vague one causes drift) and success criteria (missing them means the agent never knows when it's done) — they bracket the whole task.
2An agent touched a file it shouldn't have and ballooned the scope. Which brief part was missing?›
Constraints. Over-reach is the named failure of missing boundaries. The fix is to state what not to do and cap the scope (e.g., "don't modify the raw data; under 400 words; no external calls"), not to find a smarter model.
3Why is "make it professional" a weak instruction, and what's the fix?›
It's an adjective — it describes a feeling, and your sense of "professional" won't match the agent's. The fix is to show, don't tell: paste an example, a format, or a file to mimic. One concrete example removes the ambiguity that a paragraph of adjectives can't.
4Is a longer brief always a better brief? Explain.›
No. More words aren't more clarity — padding buries the important instructions and spends the agent's finite attention on filler. The goal is right-sizing: the shortest brief that leaves no room for the wrong guess. Cut any sentence the agent could succeed without; keep anything whose absence would let it guess wrong.