Most AI marketing assets fail before a single pixel is generated, because nobody told the model who the buyer is or what the product promises them. The tooling is almost never the bottleneck. You can wire the strongest image model available into your workflow and still get output that looks like every other AI ad.
This is the setup in the order I'd build it. First the business context, written down. Then a project that remembers it. Then a media generator driven from inside that project rather than clicked by hand. Then the loop that turns a good one-off into a format you can run again next month.
Pain, person, promise: the part everyone skips
Before you generate a single image, write down three things. The pain your product removes. The person who actually feels that pain. The promise your product makes to that person.
Keep it plain. Suppose, hypothetically, the product is a canned coffee with protein in it. The person is a busy professional who runs on coffee, cares about protein, and rushes through breakfast most mornings. The pain is that the morning currently takes three separate purchases: the coffee, the shake, and something portable to eat. The promise is that one can covers the coffee ritual and the protein at once, so leaving the house gets simpler.
A paragraph each is enough. No positioning jargon, no mission statement. What matters is that a model reading those three paragraphs can tell whether a headline is on target or off it.
This is the actual prerequisite, and the step people skip, because it feels like homework rather than magic. Skip it and you get output with no point of view: clean images that could belong to anyone selling anything. Intent has to come from you, in writing, before any generator is involved.
Why the generator sits behind Claude Code
You could open an image tool and prompt it directly. The reason I run it through Claude Code instead comes down to what survives between sessions.
A project is a folder on your machine that Claude works out of. Inside it sits an instruction file, CLAUDE.md, plus whatever context files you build up. Every new conversation starts already knowing the business. Prompt the image tool directly and you carry that context in your head, retyping it each session and forgetting a third of it.
Skills are the second reason. A skill is a saved recipe: a plain text file describing how one kind of job gets done. Install one in the project and an Instagram carousel comes out the same way every time, in the same shape, with the same slide logic. That is what makes a format repeatable instead of a lucky prompt you cannot reproduce.
Brand guidelines are the third. Drop the guidelines document into the project and every asset follows the palette, the typography and the logo rules without you restating them in each brief.
Setting up the project
Start with an empty folder, somewhere you will find it again.
Install the Claude Code desktop app, sign in, and start a new chat. Down in the corner it will say the session is running locally, on your own machine. Click to open a folder, pick the one you just made, and trust the workspace. Trusting it lets Claude read files there, create new ones and move things around. Despite the name, none of this requires you to write code.
Then ask, in ordinary language, for the scaffolding: "I run a business called X. It does Y. I'm going to paste in a pile of context about it. Please create a CLAUDE.md for this project plus any other context files you think are worth having, because you're going to act as my marketing team: ads, a website, social content." Then paste the context. The three Ps first, then anything else you have: product facts, the angles you already use, how you talk about the thing.
What comes back is a small library. CLAUDE.md becomes the operating document: it tells any future session what this project is and which files to read before writing anything. Alongside it you get context files for brand positioning, the customer avatar, product facts, voice and copy, a message bank of usable lines, playbooks, and the competitive picture. It also creates an assets folder for the material you supply and an output folder for everything it produces. Treat CLAUDE.md as a living document: when you learn something that changes how the marketing should work, say "add that to CLAUDE.md" and it goes in.
If you sit down to paste in context and find you do not know what to say, there is a better move than staring at the cursor. Use an interview-style skill: a saved prompt whose only job is to question you about one topic at a time and keep digging until it has the full picture, then write your answers into a project file. You get the skill as a text file, drag it into the chat, say "install this skill in the project", then run it. It interviews you, you answer in your own words, and the transcript becomes context the project keeps. That is the fastest way I know to move what is in your head into a file a model can read.
Next, give it real material: the logo files in light and dark versions, product photography, and the brand guidelines document. Drag them into the chat window, PDF included.
Then verify it can actually see them, because this is where people lose an afternoon. A file appearing in the sidebar is not proof the model read it. Sidebars lag, and a nested folder can look broken while being perfectly readable. So ask: "there should be 11 product shots in that folder, can you see all of them?" and have it list them back. Ten seconds of checking tells you whether the next hour of generation runs on your assets or on the model's imagination.
Connecting the media tool, then building the first thing
The media generator in this setup is an aggregator: one subscription that puts many image and video models behind a single account, plus a marketing studio that assembles finished ad creatives and music-driven reels rather than raw clips. Higgsfield is the one I am describing here. Convenience is its selling point, not price, and I will come back to that.
Connecting it takes a minute. In the tool, find the MCP and CLI section and copy the connector URL for Claude. In Claude, open Customize, go to Connectors, add a custom connector, paste the URL, name it, and add it. A sign-in prompt appears, you log in, and you grant the two services permission to talk. MCP deserves one sentence of explanation: it is a standard socket that lets an assistant operate another service directly.
After that you never navigate the generator's interface again. You ask Claude for an image and Claude goes and makes it there. You ask for a reel and it opens the marketing studio on your behalf.
For a first build I would do the website, because it exercises everything at once. There are public GitHub repositories that produce scroll-driven animated sites, code anyone can reuse. Copy the repository address, paste it into the chat, and brief the job: set up a site for the business, you already have the logo and guidelines and product shots, use the repository as the technical base, and read the three Ps and the persona so the copy and the page structure speak to that person and the pain being solved.
Two details in what came back are worth studying, because both trace back to the first section. It pulled the design system out of the guidelines document instead of inventing one, which is why the colors and the logo treatment look like the brand rather than a template. And it opened the page on the pain rather than the product: the first line addresses the reader's messy morning, and the product arrives as the answer a few scrolls later. Neither is a design trick. Both follow from the context file it read first.
Now the honest part. The site runs on localhost, meaning it exists only on your computer: send that address to a colleague and they see nothing. Getting it publicly hosted is a separate, short step you still have to do. Some sections come back blank, like nutrition facts you never supplied, because the model will not invent facts it was not given. And a first pass is a first pass.
Four formats in one chat
You do not run these jobs one at a time. You fire several briefs into the same conversation and let them work in parallel. Four of them, run at the same time, produced this.
- Static ad creatives for a buy-one-get-one promotion: a batch of around 18 images, with the product renders staying consistent across all of them, brand colors and typography applied throughout, and several deliberately left blank so copy can be dropped in later.
- A music-driven sizzle reel out of the marketing studio, delivered in both vertical and landscape cuts, with scenes, music and on-screen copy already assembled.
- Five Instagram carousels of seven to eight slides each, built around different angles: why the product exists, who it is for, what problem it removes, the flavor line-up.
- Vertical UGC-style video ads, which it approached in stages: first a cast of four characters as still images, then storyboards built from those characters, then video generated from the storyboards, then its own quality pass, taking screenshots of each take and rejecting the ones where the packaging came out wrong.
That last sequence is the interesting one. Generating characters before storyboards, and storyboards before video, is what keeps a clip coherent instead of drifting between shots. It also means the cast is reusable: the same four faces can front next month's campaign.
The prompt shape that makes all of this work is not clever. It has four parts: the motivation, the end goal, where to file the output, and a request for the cost. "We have a promotion coming up, here is the offer, I want a spread of creatives to test, use our brand context, save them in the output folder under this campaign name, and tell me what it cost." Plain language, the way you would brief a person who already knows the business, because at this point it does.
Ask about cost every time and you build a real sense of what each format is worth. The pattern holds: a batch of static creatives runs to a few dollars, carousels sit a little above that, and a finished music-driven video costs several times more than either. The staged UGC videos are the most expensive line, because each one pays for images, storyboards, video generation and a quality pass. Knowing those ratios is what lets you decide which format to run at volume.
Turning one-offs into a system
The part that makes this accumulate is a habit rather than a feature.
When something lands, say so and turn it into a skill. "I like this carousel style, make a skill out of it so I can produce that look again." When something misses, say that too: "never produce them like that again." A skill saved from a result you liked is the difference between a lucky output and a format you own. Do that a few times and the project stops being a generator you prompt from scratch and becomes a set of recipes you call by name.
The reusable pieces work the same way. The cast of four characters is an asset, not a by-product. So are the creatives left blank on purpose: the next promotion swaps the headline instead of regenerating every image, which is faster and cheaper.
Then close the loop with a tracker. Ask for a spreadsheet and describe the tabs you want: a generation log with the date, the type of run, the campaign, the channel, the primary angle, the deliverables, which model produced them, what it cost and the status. A sheet listing the creative assets with their copy. A tab for angle performance. A copy bank of lines worth reusing. A cost summary.
The tracker is worthless until you feed it real numbers. Once performance data goes in, views, engagement, conversions, you can hand the whole sheet back and ask for a weekly review: what worked, what did not, make more of the first kind. That review is the only mechanism here that improves the output rather than multiplying it. Volume buys you attempts and keeps you from betting everything on one guess. Measurement tells you which attempt to repeat. Neither does much without the other.
Where the output still falls short
The UGC videos still read as advertising. They are watchable and the cuts are quick enough to hold attention, but nothing about them feels like a person who happened to film their morning. The fix is not a better model. It is having the project study genuinely organic content first, then write scripts you read and approve before a single frame gets generated.
No subject-matter expertise went into any of the prompts I described. The brief for 18 ad creatives amounted to "make me some ads for this promotion." If you have run paid social, if you know which hooks survive contact with a cold audience, that knowledge in your prompt is worth more than any model upgrade on offer.
The aggregator is convenient rather than cheapest. One subscription across many models saves you accounts and API keys, and you pay for that in unit cost. Worth it while you are learning what you need, worth re-examining once you know which two models you actually use.
And one prompt is not a launch. What this setup gets you is a standing ability to produce branded assets on demand, at a cost low enough that testing stops being a budget conversation.
The asset generation was never the scarce part, and it gets cheaper every quarter. The scarce part is knowing who you are talking to and what you are promising them, written down where a model can read it. So write the three paragraphs first. Everything after that is plumbing you can assemble in an afternoon.