STOP PROMPTING, START BRIEFING: CLAUDE COWORK FOR DESIGNERS
The tech industry has spent two years teaching itself something every design studio already knew. They gave it a name, context engineering. There are job titles now, conference talks, an entire cottage industry of prompt packs. Strip the vocabulary off and what's underneath is a creative brief.
That should annoy you a little. It should also make you feel great, because it means the skill everyone is scrambling to learn is one you've had since design school.
What an AI agent is, minus the jargon
An AI agent is AI that works like a junior designer instead of a vending machine.
A vending machine gives you one output, once, and if it's wrong, tough luck. That's a prompt. An agent works in a loop: it does the work, looks at the result, decides it's not right yet, and goes again until the job is done. Simon Willison, one of the most trusted voices in this space, defines it in eleven words: an agent "runs tools in a loop to achieve a goal."
Read that back slowly. Does the work. Checks it against the goal. Revises. You've sat in that room a thousand times. It's a junior with a brief, and the loop is just iteration, the thing we've always known good work requires.
Which changes where the skill lives. A prompt rewards clever phrasing. A loop rewards knowing what "done" looks like before you start, what the constraints are, who the audience is, what you never want to see. Designers have a document for that.
If you've played with node-based tools like Weavy, you've already met the ancestor of this idea: prompts turned into repeatable workflows. Agents take the next step. The workflow builds itself; you supply the judgement.
I gave it one messy client call
Claims are cheap, so I tested it. Meet Frankie: he's opening a late-night gelato spot on Gertrude Street in Fitzroy, open till 1am, working name Scoop Therapy. Investor dinner in five weeks. Frankie is fictional. Everything else about this process is real.
I gave Claude Cowork two files: a messy 27-minute discovery call transcript, and a one-page brief written the way I'd brief a junior strategist. The prompt itself was two sentences, because the thinking lives in the brief, not the chat box. Nobody ever art-directed a campaign from a Post-it note.
Here's what came back, and where it got interesting.
It pitched me a plan before touching anything. Because the brief told it to. I stayed the creative director; it stayed the junior.
It fact-checked my client. Frankie claimed a competitor closes at 8:30pm. The agent went and verified every market claim online, with sources, and found they push to 10pm on Fridays. It flagged the correction before an investor could. How many decks have you seen where a client's throwaway claim went straight onto the slide?
It ran a crit on its own first draft. The brief required a draft file, then a critique against the brief, then a revised final. The crit found a real contradiction: the strategy claimed an "empty" late-night market while its own competitor table showed one venue open till midnight. It caught that and rewrote the claim into something defensible.
Then I tried to break it. The work came back good, which is exactly when a creative director should get suspicious. So I picked my favourite territory, a warm-glow-in-the-window idea, and asked it to stress test the concept from the perspective of the actual audience. It researched the real venues on Gertrude Street and came back with the kill shot: every bar on that strip already owns warm light. Its line, and I'm still not over it: "Owning warm light on Gertrude Street is like owning exposed brick in Collingwood." Then it rebuilt the territory around domestic light instead of venue light. A house lights its window; a bar lights its room.
Twelve minutes of working time turned a transcript into a positioning map, a strategy one-pager and three distinct territory briefs. Not finished work. A really good first crit round.
The six-part agent brief
This is the framework, and you already know it, because it's the anatomy of every decent creative brief ever written:
Deliverable. What you want, in what format, in what order.
Audience. Who reads it and what they need to feel. Investors read differently to design teams. Say so.
Constraints. Budget reality, scope boundaries, what's out of bounds.
References. The source material, and which language in it matters.
Definition of done. The self-crit. Make it check its own work against named criteria before handing anything in, and make it show the critique.
Anti-goals. What you never want to see. Ours included "no fabricated market data" and my favourite line in the whole document: don't sell the gelato, sell the hour.
One distinction worth stealing: split your missing facts into two kinds. World facts (competitor hours, name clashes, market claims) the agent should verify online and cite. Client facts (budget, seat count, the co-founder's opinion) it can't look up, so it should flag them as open questions instead of inventing answers. That single instruction is the difference between a research partner and a confident liar.
Claude or ChatGPT?
We've run this comparison before in sillier circumstances, when I asked ChatGPT and Claude to roast each other. For agentic work, the honest answer: they're converging, and the brief transfers between them. Where they differ today:
| Claude Cowork | ChatGPT Agent mode | |
|---|---|---|
| Built for | Working in your files and folders, delivering finished documents | Web-native tasks: browsing, forms, research |
| Designer fit | Brand audits, strategy docs, multi-file studio jobs | Competitor monitoring, sourcing, admin |
| The loop | Deep, persistent, file-based; the draft, crit, final pattern | Shorter sessions (roughly 5 to 30 minutes), easy recurring schedules |
| Watch-outs | Desktop app, paid plans | Session caps; parts of its builder tooling are being retired |
For studio work of the kind in this demo, Cowork's file-based loop is the better fit. For "check what our competitors post every Monday," ChatGPT's scheduling is friendlier. Neither one rescues a vague brief.
Taste doesn't loop
Now the caveats, because this is where the whole thing either becomes a practice or a liability.
The agent's output was impressive and it was not client-ready. Which territory actually goes in front of Frankie's investors is a taste call, and taste doesn't loop. The machine did the week-one grind: the reading, the mapping, the first structured thinking. Choosing, curating and directing stayed human, and always will, because that's the scarce part.
And you own what your agent ships. If it asserts a competitor's closing time, you check the citation. If it writes strategy from a client call, you confirm nothing was invented. The crit pass in the brief isn't a party trick; it's the ethics of the workflow. An agent without a definition of done is just AI slop with more steps.
"The prompt is not the art. The creative direction is the art."
That was true of image generators and it's truer of agents. A vague brief gets you slop at scale, the same way a vague client brief gets you a logo that "needs to feel premium but approachable." Garbage in, garbage out didn't stop being the rule. It just got faster.
Your move
Take a job you'd hand a capable junior: a competitive audit, a research summary, the first pass of a strategy doc. Write the six parts. Give it to an agent. Then do what you'd do with any junior's work: put it on the wall and crit it.
Still building the foundations? Start with how to integrate AI into a real design workflow, then come back and hand your first brief to an agent.
While everyone else is buying prompt packs, you've been holding the actual skill since design school.
Stop prompting. Start briefing.
Want to see the full run, including the moment it tried to kill my favourite idea? Watch the demo above, and if you want to go deeper, this is the kind of workflow we build together inside the Creative Futures Hub. And if you're ready to learn how to integrate AI into your creative process properly, from strategy through to identity, that's what the AI Branding Masterclass is built for.
FAQ
What is an AI agent? An AI agent is a language model that runs in a loop: it takes an action, checks the result against a goal, and repeats until the job is done. For designers, think of it as AI that works like a junior designer with a brief rather than a vending machine that gives you one output per prompt.
What's the difference between a prompt and an agent? A prompt is a one-shot instruction with one output. An agent works iteratively: planning, doing, self-checking and revising across many steps, and using tools like file access and web search along the way. Prompts reward clever phrasing; agents reward a clear brief.
Is Claude or ChatGPT better for design studio work? Claude Cowork is stronger for file-based studio jobs like brand audits and strategy documents, because it works in your folders and delivers finished files. ChatGPT Agent mode is stronger for web-based tasks and recurring scheduled checks. The briefing skill transfers between both.
How do I brief an AI agent? Use the same anatomy as a creative brief: deliverable, audience, constraints, references, a definition of done that forces a self-critique, and anti-goals (what you never want to see). Tell it to verify world facts online with sources and to flag client facts it can't know as open questions.