Building an AI-powered brief generator for client projects

An AI brief generator sounds like a small quality-of-life improvement. In practice, it changed how we start every client project. At Choco Media, we spent years watching good work slow down at the intake stage — not because clients were difficult, but because the information we needed to do strong work was scattered, vague, or missing entirely. A thirty-minute briefing call would end and we’d still need to follow up three times before we could open a document and actually start. When we replaced our manual intake process with a GPT prompt chain, that friction nearly disappeared. This post documents exactly how we built it, what it takes in, and what it produces.

This is written for marketing teams, agency leads, and in-house brand managers who run a steady volume of projects — content, campaigns, ads, brand work — and find themselves spending more time chasing context than creating. You don’t need a developer to build what we describe here. You need a structured intake form, a well-designed prompt, and a way to stitch them together. We’ll walk through all three.

By the end you’ll have a clear picture of the intake questions that actually matter, the prompt architecture that transforms raw answers into a usable brief, and the checks we built in to stop AI-generated briefs from hallucinating client intent.

Why traditional client briefs fail before you write a single word

Most brief templates were designed for a slower era. They ask for things like “campaign objectives” and “target audience” without providing any scaffolding for how to answer those questions — so clients fill them in at whatever level of depth they feel comfortable with, which is almost never the level you actually need.

The result is a brief that looks complete but isn’t. It has words in every field. It has a deadline and a budget. What it doesn’t have is the specific tension the campaign needs to address, the one objection the audience carries, or the tone boundary that separates their brand from every competitor. Those are the things that make creative work specific rather than generic, and they almost never appear in a client’s first pass.

• Open-ended questions produce open-ended answers. “Describe your target audience” gets you “professionals aged 25–45 who care about quality.” That’s a demographic, not an insight.
• Clients underestimate what they know. They have strong opinions about what feels wrong for their brand, but no one has asked them to name it explicitly.
• Follow-up emails are expensive. Each clarification round adds a day or more to a project and trains clients to expect delays.
• Brief quality predicts output quality. Every revision cycle we’ve traced back to a vague brief. Not most — every one.

An AI brief generator doesn’t fix bad questions. It makes the intake questions better, then turns complete answers into a structured document without the interpretation layer that causes misalignment. That’s the leverage.

What goes into the intake form

The intake form is the foundation. If you rush it, the AI has nothing to work with and you’re back to follow-up emails. We spent a month refining ours, cutting it from 22 questions to 11, all of them closed or semi-closed so clients can answer quickly but still provide the specificity we need.

The questions we kept

• What does this project need to make true for your audience? Not “what’s the goal” — what belief, feeling, or decision needs to shift.
• What’s the one thing your audience currently thinks that this project needs to change or reinforce? Forces the client to name the before-state.
• What’s off-limits — in tone, visuals, or messaging? This single question has saved more revision rounds than any other.
• Who is this NOT for? Exclusion criteria sharpen targeting faster than inclusion criteria.
• If this project fails, what probably went wrong? Surfaces the client’s hidden anxieties and gives you their definition of risk.
• Which existing piece of your work (yours or a competitor’s) comes closest to what you want? A reference is worth five paragraphs of description.

The other five questions cover practical scope: deliverables, deadline, approvers, usage rights, and budget range. Those are necessary but not where the intelligence lives.

Designing the prompt chain

A single prompt won’t produce a complete brief. We use a three-stage chain: a parsing pass, an expansion pass, and a formatting pass. Each stage has a specific job and passes its output to the next.

Stage 1: Parse and check for gaps

The first prompt reads all intake answers and does two things: it extracts the core inputs into a structured JSON object, and it flags any answers that are too vague to work with. We tell the model explicitly to return a “gap_flags” array with any field where the answer won’t support strong creative direction. If three or more flags appear, we trigger a short follow-up email — automated, targeted, and limited to the specific gaps.

• Input: raw form answers
• Output: structured JSON + gap_flags
• Model: GPT-4o (we’ve also used Claude 3.5 Sonnet for this stage with similar results)

Stage 2: Expand into brief sections

The second prompt takes the structured JSON and writes five brief sections: strategic context, audience insight, creative territory, mandatory constraints, and success signals. Each section has a prescribed length — usually two to four sentences — so the output is dense rather than padded.

This is where most of the quality lives. The prompt for this stage is the one we’ve iterated most. The key instruction is a negative one: do not invent specifics that weren’t present in the intake answers. The model should synthesise and articulate what the client said, not add colour it imagined. We test this regularly by running briefs where we’ve deliberately removed details, and checking whether the model invents plausible-sounding substitutes. It does — that’s what it’s trained to do — so the instruction needs to be explicit and the output needs a human check.

“The prompt that produces the best brief isn’t the cleverest one — it’s the one that knows what not to invent.”

Stage 3: Format for humans

The third prompt takes the five sections and renders them as a clean Markdown document with a header, section titles, and a one-paragraph summary at the top. This is the brief the team actually works from. We pipe it into Notion automatically; some teams prefer to export it as a PDF for client sign-off.

The stack we use (and what it costs)

We built our version in n8n for orchestration, Tally for the intake form, and the OpenAI API for the model calls. Total monthly cost for the API calls runs under €30 for a volume of 20–30 briefs. The Tally free plan covers most team sizes; n8n self-hosted is free.

• Tally — free tier, handles conditional logic, exports to JSON or webhooks
• n8n — orchestration; one workflow with three HTTP request nodes for the three prompt stages
• OpenAI API — GPT-4o for stages 1 and 2; GPT-4o-mini for the formatting pass (cheaper, no quality difference for formatting)
• Notion — final destination; the brief appears as a new page in a “Briefs” database automatically

If you’re not comfortable with n8n, the same chain runs in Make (Integromat) with near-identical setup. We cover this kind of AI automation build in our service work — if you want the workflow built rather than DIY’d, that’s an option too.

What the output actually looks like

A completed brief from the chain runs about 400–500 words. It’s shorter than a traditional brief template, but denser. Every sentence is load-bearing. The strategic context section names the specific tension the project addresses. The audience insight section describes the audience’s current state and the shift the project needs to create. The creative territory section gives the team two or three directions to explore, not one predetermined execution.

The mandatory constraints section is often the most valuable. It captures the client’s explicit limits — things like “never use the word ‘solution'” or “no stock imagery of people in suits” — alongside the implicit ones the intake questions surfaced. In client work we’ve found that constraints section alone accounts for a significant share of first-draft acceptances, because it prevents the team from producing work the client will reflexively reject before even reading it.

• Strategic context: 2–3 sentences on why this project exists and what it needs to shift
• Audience insight: who they are, what they currently believe, what needs to change
• Creative territory: 2–3 directions, not executions
• Mandatory constraints: explicit and implicit limits on tone, visuals, and messaging
• Success signals: how the client and team will know the project worked

The human check that keeps the system honest

We run a two-minute review on every generated brief before it goes to the team. This isn’t about correcting the AI’s writing — the language is usually fine. It’s about catching hallucinated specifics and flagging briefs where the intake answers were thin enough that the model had to fill gaps it shouldn’t have.

Our review checklist has four items:

1. Does every specific claim in the brief trace back to something the client actually said?
2. Is the audience description grounded in the intake, or does it sound generic?
3. Are the creative directions genuinely distinct, or are they the same idea in different language?
4. Would the client recognise their project in this brief, or would it feel like it could belong to anyone?

If the answer to any of these is no, the brief goes back to the gap-flagging stage and we follow up with the client. This happens on roughly one in six briefs — usually because the original intake answers were short. The system doesn’t replace judgement; it handles the work that doesn’t require it.

Time savings and what we do with them

Before the AI brief generator, our briefing process took 45 minutes on average: a 30-minute intake call, plus 15 minutes of document production. Now it takes 10–12 minutes: five to seven minutes for the client to complete the form, two minutes for the chain to run, and two minutes for the human review. That’s roughly 33 minutes recovered per brief.

At 25 briefs a month, that’s about 14 hours — more than two full working days — returned to the work itself. We use that time upstream (more thorough content strategy and production work) and downstream (more iterations on early creative, rather than starting from scratch after a misaligned first draft).

In client work we’ve also found that the intake form itself improves the quality of thinking clients bring to the project start. The questions force them to articulate things they’d previously left implicit. By the time the brief lands in Notion, the client has already done more strategic thinking than they would have in a traditional 30-minute call. That clarity carries through the whole project.

How to start if you’re building this from scratch

The fastest path is to take your existing brief template, strip it to the six to eight questions that actually affect creative output, and build a Tally form around those questions. Then write a single prompt — not a chain — that takes the answers and produces a first-draft brief. Run that for a month before adding stages two and three. The chain adds precision, but the single-prompt version is already faster and more consistent than a manual process.

• Start with one prompt, not a chain
• Use your worst recent brief as the test case — if the AI improves on it, you’re on the right track
• Add the gap-flagging stage once you’ve identified which questions your clients consistently answer poorly
• Build the Notion integration last — email delivery works fine while you’re calibrating the prompts

If you’d like to talk through whether this kind of build makes sense for your team’s workflow, we’re easy to reach — we usually start with a short conversation before anything else.