How to Brief an AI Agent: The Prompt Patterns That Actually Work
To brief an AI agent effectively in 2026, give it four things: the goal (what success looks like), the context (who you are and who you are talking to), the constraints (tone, length, format, what to avoid), and a worked example or reference (what good looks like). The single biggest mistake is asking for output without specifying success criteria. An agent given "write a blog post about marketing" produces generic copy. An agent given a goal, context, constraints, and example produces work that is directly usable. The pattern is identical to briefing a freelancer, except agents can absorb 10 times more context per brief and never charge by the hour. This guide walks through the four parts, gives you a copy-and-paste template, and shows the brief patterns that work for the most common marketing tasks.
TL;DR
- The four-part brief is goal, context, constraints, reference
- Goal is the most-skipped part and the highest-leverage to get right
- Context replaces every generic output with a specific output
- Constraints prevent the agent from wandering off the requested format
- A reference example is worth more than three paragraphs of instruction
- The brief template below works for any marketing task
The four parts of a brief that works
A weak brief is one sentence: "Write a blog post about AI marketing." A strong brief is structured. The four-part shape below is what works.
| Part | Question it answers | What it looks like |
|---|---|---|
| Goal | What does success look like? | A measurable outcome, a target reader response, or an explicit definition of done |
| Context | What does the agent need to know? | Company, audience, voice, constraints of the world this work lives in |
| Constraints | What rules apply? | Length, format, structure, words to avoid, tone, output type |
| Reference | What does good look like? | A previous winning piece, a competitor example, or a structural template |
The four parts compound. Each one alone improves output by 20 to 30%. Together they improve output by an order of magnitude.
The brief template
This is the template you can copy into any AI agent or chatbot.
Goal
What I want this output to do, who reads it, and what action they take after.
Context
About my company: [one paragraph]
About my audience: [one paragraph]
About my brand voice: [three or four characteristic words plus any vocabulary rules]
About the project: [why this output exists in the broader campaign]
Constraints
Length: [word count or token range]
Format: [structure, sections, tables, FAQ section, etc.]
Voice: [first person plural, second person, etc.]
Words to avoid: [any banned terms or phrases]
Words to use: [primary and secondary keywords if relevant]
Output style: [markdown, JSON, plain text]
Reference
A previous output that worked: [paste the text or link]
What about it worked: [one or two sentences]
What I want different this time: [one or two sentences]
That template is 12 lines of structure. Filling it in for a typical task takes 10 minutes. The output is consistently 5 to 10 times better than a one-line prompt.
Brief patterns by task type
Different task types need slightly different emphasis in the brief. The table below shows where to put the weight.
| Task type | Strongest part | Notes |
|---|---|---|
| Blog post | Reference (previous winning post) | The structure is the work |
| Cold email | Constraints (length, voice, CTA) | Tight format wins replies |
| Sales sequence | Goal (specific deal stage) | Stage-aware messaging is the difference |
| Case study | Context (the customer story) | Without raw material it cannot write |
| Ad copy | Constraints (character limits, ICP) | Format discipline wins tests |
| Battle card | Reference (competitor positioning) | Comparison is the deliverable |
| Web page | Reference (similar pages that work) | Structure matters more than copy |
| Email nurture | Goal (where in the journey) | Wrong stage equals wrong tone |
Use the same four-part template for all of them. Just emphasise the right part.
A worked example
Below is a real brief for an Orbitable agent producing a blog post. This level of detail is the norm, not the exception, in production workflows.
Goal
Produce a 2,000-word blog post optimised for the keyword "AI marketing for mid-market B2B SaaS" that ranks in Google's top 5 within 90 days and gets cited by ChatGPT and Perplexity when users ask about AI marketing for SaaS companies.
Context
About the company: Orbitable is an AI marketing platform with 53 specialist agents that handle the full GTM lifecycle.
About the audience: B2B SaaS marketing leaders at companies between 50 and 500 employees who are evaluating AI tools and trying to understand fit for their stage.
About brand voice: Direct, evidence-led, slightly technical, never hyped. We do not use phrases like "AI-powered" or "transformative". We do use specific numbers and concrete examples.
About the project: Part of a series targeting mid-market B2B SaaS audiences. Should sit alongside the playbook post and the multi-agent explainer.
Constraints
Length: 1,800 to 2,200 words
Format: H1 title, direct answer paragraph, TL;DR, multiple H2 sections, at least three data tables, FAQ section with 5-6 question-answer pairs
Voice: First person plural ("we") sparingly, mostly third person
Words to avoid: revolutionary, transformative, AI-powered, game-changer, solution
Words to use naturally: AI marketing, B2B SaaS, multi-agent, world model, ICP, ABM
Output style: Markdown, British spelling, no em dashes (use periods, commas, conjunctions)
Reference
A previous post that worked: [paste of the AI Marketing for B2B SaaS Playbook post]
What worked about it: Tables broke up the prose, the four-phase structure mirrored how teams actually think, and the FAQ caught long-tail searches.
What different this time: Tighter on tools and deeper on the metrics. The previous post was strategic. This one needs to be practical.
That brief is 350 words. The output is 2,000 words of production-ready content. The ratio of brief work to output work is the inverse of what most teams do, and it is exactly why their output is generic.
What goes wrong when briefs are weak
The patterns are predictable.
No goal: Output is generic and could apply to any company in the category. The reader cannot tell what to do next.
No context: Output reads like it was written by an outsider who looked you up on Wikipedia. Specific terms are missing or wrong.
No constraints: Output drifts in length, tone, or format. You spend the same amount of time editing as you would have writing it yourself.
No reference: Output uses a structure that is plausibly correct but wrong for your audience. Tables appear where prose was wanted, prose appears where lists were wanted.
The brief is the lever. Almost every "AI output is bad" problem is a brief problem.
Iterating on a brief that almost works
When the first output is close but not right, the iteration pattern is more efficient than rewriting from scratch.
- Identify the specific gap (length, tone, missing section, wrong angle)
- Add one constraint or one paragraph of context to address it
- Re-run the brief
- Repeat until the output needs only line edits
Most briefs reach production quality in three to five iterations. That is 30 to 50 minutes of work to lock a workflow that runs forever.
The classic mistake is to bin the output and start over. The agent had most of it right. Find the gap, patch the brief, run again.
How briefs differ between single agents and multi-agent fleets
If you are using a multi-agent fleet with a shared world model, the context block of the brief shrinks dramatically. The world model already holds the company, audience, and voice. Your brief becomes mostly goal, constraints, and reference.
If you are using a single agent without persistent state, the context block stays full every time. Same brief, every prompt. This is the structural reason fleets compound: the context block is reusable.
| Single agent | Multi-agent fleet | |
|---|---|---|
| Brief size per task | 300-500 words | 100-200 words |
| Re-briefing required | Every conversation | Once per project |
| Brief reusability | Manual copy-paste | Saved as template |
| Brief drift over time | High (copy-paste errors) | Low (single source) |
This is one of the unsung advantages of fleets. Less brief, less drift, more leverage.
FAQ
How long should an AI agent brief be?
For a single agent without persistent context, a strong brief is 300 to 500 words. For a multi-agent fleet with a shared world model, 100 to 200 words is enough because context is already loaded. The shorter brief on a fleet still produces better output.
Should I write the brief before or after I have an output to react to?
Always before. Briefs written after the fact tend to inherit assumptions from the bad output. Writing the brief first forces clear thinking about success criteria, which is the highest-leverage step.
What if the agent ignores parts of my brief?
This usually means the brief is too long or contradictions exist between sections. Tighten the wording, remove redundancy, and check that goal, context, constraints, and reference all point in the same direction.
Do I need to use specific prompt engineering techniques like chain-of-thought?
For most marketing tasks, no. The four-part brief is more important than any specific prompting technique. Chain-of-thought helps for analytical tasks (scoring, ranking, decision-making) but rarely for creative production.
How do I brief multiple agents on the same campaign?
Brief the campaign once at the world-model level (audience, voice, segment), then brief each agent with the goal and reference for its specific task. The shared context inherits, the per-task brief layers on top.
What is the most-skipped part of the brief?
Reference. Most teams write goal, context, and constraints but skip the example. Adding a single worked example to a brief is usually the highest-impact change available.