Brief Writing with AI

The magic isn’t in the prompt. It’s in the brief you translated into one.

The freelancer who ships great AI-assisted work first-pass isn’t doing magic prompting. She’s writing a better brief, then translating it into a prompt the model can act on. Watch anyone share their “prompt secret” and look closely: most of what they call prompt-craft is brief-craft. The AI part is the easy part.

Here’s the thesis. A creative brief written in plain English is a contract with yourself. The same brief written as a structured prompt is a contract with the model. The translation between them separates usable AI work from generic AI work. It’s a skill, not a trick, and this article teaches it.

This is the second Workflow anchor, after Build Taste, Generate, Refine. It sits at Phase 2 (Generate), on the most consequential pre-step: deciding what to generate. Phase 1 is Moodboarding with AI; Phase 3 is Design Critique with AI. Brief-writing is the bridge.

It’s also the step most AI workflows skip. The branded, multi-stage workflows aimed at seasoned designers tend to start at discovery and assume the brief already exists; they teach what to do with a brief, not how to write one. That gap is what this article fills.

A brief is a contract with yourself: what you’re making, for whom, under what constraints, in what voice. A prompt is a contract with the model: what you’re asking it to produce, with what constraints, in what format. Same function, two layers, two languages.

Most “AI for designers” content treats the brief and the prompt as one artefact. You type the brief into the tool and get generic output back. The brief asked for “modern, minimalist, premium,” three median descriptors, so the model produced the median. It did exactly what you said.

The move is translation. Take the brief’s plain-English requirements and turn them into model-actionable constraints. Modern becomes “sans-serif geometric grotesque, lowercase, generous tracking.” Minimalist becomes “60% blank canvas, single focal element, restrained accent palette.” Premium becomes “high figure-ground contrast, restrained saturation, no decorative chrome.” The brief said three words; the prompt says fifteen. Same client requirements, different language, usable output.

Design writers are converging on this: Smashing Magazine calls prompting “a design act” and frames a good prompt as a creative brief for a non-human collaborator. The skill you have transfers; you just aim it at the model.

To make that translation repeatable, we use a four-part pattern: Role, Context, Task, Tone, or RCTT for short. It isn’t a published framework; it’s a working synthesis, and we’ll name it as one. But the four pieces aren’t ours to invent. Every major model-maker’s guidance recommends them: Google, OpenAI, and Anthropic all tell you to set a role, give context, state the task, and steer the tone. Singapore’s GovTech built context and tone into its CO-STAR pattern; the common RTF pattern carries role and task. It bundles the four that matter most for a brief. Operationally useful in practice, not a benchmark.

Role. Who is the model being? “Senior brand designer,” “junior copy editor,” “creative director reviewing for accessibility.” The role narrows output toward what that role would produce. Without one, the model defaults to generic-helpful, which is what produces median work.

Context. What does it need to know? Audience, brand voice, the surrounding work, accessibility constraints, what’s already been ruled out. Context is the dimension that moves the output most. Most brief-to-prompt failures are context failures.

Task. What specifically should it produce? Format (deck slide, hero image, 30-second tagline), scope (one variation or a batch of eight), and success criteria (passes WCAG AA). Without a tight task the model either under-delivers or hands back everything plus what you didn’t ask for.

Tone. What voice should it carry? “Calm-confident-not-loud,” “warm-but-disciplined,” “editorial-restrained.” Tone is what makes the output feel yours rather than the model’s median.

Each one closes a category of drift. Role closes generic helpfulness, Context closes audience-blindness, Task closes scope drift, Tone closes voice drift. Together they shape the output before the “generate” line ever fires.

RCTT (Role / Context / Task / Tone: the four dimensions a brief-to-prompt translation should specify; a TGDS working pattern, not a measured benchmark). Brief-to-prompt translation (moving from plain-English requirements to model-actionable constraints). Constraint set (the bundle of constraints a prompt carries; constraint sets get honoured, wish lists don’t).

The brief. A small specialty coffee roaster needs a brand mark and a one-page e-commerce hero for a 2026 launch. Audience: design-aware urban professionals. Voice: confident-warm-not-folksy. Accessibility: the hero must pass WCAG AA.

Brief typed in verbatim. Paste it in as written. What you’d expect back: a median coffee mark in the usual warm-orange skew (the bias Colour Theory and AI measures), a stock-illustration hero, and no relationship between the two.

Same brief, translated through RCTT.

Role: Senior brand designer reviewing your own work for an indie launch. Context: Specialty coffee roaster, founded 2026; audience is design-aware urban professionals (not “coffee enthusiasts” generically). The category defaults to ochre-orange, warm-folksy branding; this brand is confident-warm-not-folksy. The hero must pass WCAG AA against its gradient ground. Task: Three brand-mark variations in geometric grotesque (no orange-yellow skew; palette anchored at #1F2937 / #F4ECD8 / accent #C0392B), single weight, lowercase. One matching hero: brand mark dominant at the upper-third focal intersection, supporting copy at 60% scale, body copy at WCAG AA against the gradient. Tone: Editorial-restrained, not commercial-loud. Single accent reserved for the CTA. No decorative chrome.

What you’d expect back now: brand-coherent, distinctive, accessibility-passing work that looks like somebody chose the constraints. Same client requirements. Different language. Usable output.

Now hold the brief steady and change only the prompt’s shape.

Literal prompt. Concatenate the RCTT-translated brief into one prose paragraph: “As a senior brand designer for an indie specialty-coffee launch, design-aware urban audience, confident-warm-not-folksy, produce three brand-mark variations and a matching hero…” The model parses the prose and produces variations. Good, better than the verbatim brief, but not as reliable as what comes next.

Constrained prompt. Same content, restructured into a constraint set:

ROLE: Senior brand designer. CONTEXT (honour; do NOT default to median):

• Audience: design-aware urban professionals
• Voice: confident-warm-not-folksy
• Avoid: the orange-yellow skew typical of mid-2020s coffee branding
• Palette: anchored at #1F2937 / #F4ECD8 / #C0392B
• Accessibility: WCAG AA contrast on hero body copy TASK:

1. 3 brand-mark variations, geometric grotesque, single weight, lowercase, no orange skew.
2. 1 hero: brand mark dominant at the upper-third focal intersection; supporting copy at 60% scale; body copy WCAG AA against gradient. TONE: Editorial-restrained, not commercial-loud. Single accent for the CTA. No decorative chrome. NEGATIVE: No stock-illustration look. No centred symmetrical composition. No “modern” or “premium” generic descriptors.

The model honours numbered, labelled constraints; it interprets prose toward the median. That’s the point of Figma’s Cooking with constraints, whose TC-EBC structure (Task, Context, Elements, Behavior, Constraints) is a sibling to this pattern. Frame a prompt as a constraint set, not a wish list, and the output tightens.

Three failures show up again and again, and they share a shape.

Under-specified mood. “Modern, premium”: two median descriptors that translate to a generic prompt and a generic output. Fix: replace them with the compositional, typographic, and colour moves you actually want.

Over-specified colour. “Use exactly #1A8FE6, full saturation, only that.” The model treats the hex as a constraint, then reaches for the median palette around it, and the secondary elements drift back toward the orange skew. Fix: name the harmony the hex anchors (triadic, analogous, complementary, split-complementary). See Colour Theory and AI for the vocabulary.

Missing constraints. The brief says what to make but not what to avoid, so the model fills the gaps with its defaults: centred composition, stock-illustration look. Negative prompts are as load-bearing as positive ones. Say what to exclude.

All three leave the model room to default to its training-data median. Tighten the brief and you tighten the output. Brief precision is prompt precision.

By use-case, not by ranking:

• Claude. Best for structured prose; its output is easier to audit. Use it for brief drafting and structured deliverables.
• ChatGPT. Most familiar, widest ecosystem. Use it when the work touches non-design tools: image generation, web search, code.
• Gemini. Strong on long context (paste a whole brief library) with image generation built in. Use it when the brief lives in Google Workspace.

The tool isn’t the brief. Whichever model runs it, the RCTT translation and the constraint-set framing are what produce usable output.

Copy this into Notion, Obsidian, or a Markdown file and fill it in.

# Brief: [Project Name]

• Client: [name]
• Audience: [demographic + psychographic, be specific]
• Brand voice: [3 adjectives + 1 anti-adjective; e.g. “calm-confident-not-loud”]

• Palette: [hex anchors + named harmony]
• Typography: [classification + named anatomy choices]
• Accessibility: [WCAG AA / AAA target]
• Avoid: [bias directions, generic patterns, model defaults]

1. [Specific format + scope]
2. [Variation count + success criteria]

[1–2 sentences; editorial-restrained / commercial-loud / etc.]

• [Patterns to exclude]

The structure carries the prompt loop. Use it, modify it, make it yours.

A brief specifies typography (classification, anatomy, scale), colour (anchors and harmony), hierarchy (focal points and level counts), and composition (thirds, negative space, figure-ground). Those are the four design fundamentals under different names. The brief is where fundamentals enter the AI loop.

It maps to the phases, too. Phase 1 (Build Taste) supplies the references that fill the brief’s Constraints section. Phase 2 (Generate) is what the brief is for. Phase 3 (Refine) inherits the brief’s success criteria for the critique loop. The thesis collapses to one line: AI multiplies the brief. Feed it vague requirements and it multiplies vagueness; feed it specific ones and it multiplies specifics.

TGDS Verdict. Brief-writing is the highest-return skill a designer can sharpen in the AI era, and it’s trainable: the same vocabulary that makes a good brief makes a good prompt.

Design@Work is the applied route for working freelancers turning brief-writing into a repeatable AI workflow. Cert IV in Design is the foundational route for building the typography, colour, hierarchy, and composition vocabulary a brief draws on.

Outbound: Moodboarding with AI (the references that feed the brief’s context), Design Critique with AI (the criteria that evaluate the output), Text and Copy tools (Claude, ChatGPT, Gemini compared). For the basics, How to Use ChatGPT for Design covers the entry-level brief-to-prompt patterns; this is the structured upgrade.

Write the brief well and the prompt writes itself.