Run a one-person marketing team inside Claude Code — Checklist · Resource

Open this next to the work, not after it. The sections run in build order, and section 0 decides whether anything below it is worth generating. Every item is something you do and can then check.

0. Pain, person, promise

Write these three paragraphs before you open anything. A model that has read them can tell a headline that lands from one that misses.

The pain: ______

The person: ______

The promise: ______

[ ] Each one is a paragraph of plain language, with no positioning jargon.
[ ] The pain is what the buyer's day looks like now, in concrete steps, not a category word like "convenience".
[ ] The person is one human with a routine you could describe out loud, not a demographic bracket.
[ ] The promise names what changes for that person, stated without a single product feature.
[ ] Test the set: write two headlines, one on target and one deliberately off, and check the paragraphs are specific enough to separate them.
[ ] All three sit in a project file, not in your head.

This is the prerequisite, not the warm-up. Skipping it is why so much generated marketing looks interchangeable.

1. The project, from an empty folder

[ ] Create an empty folder you will find again, then open it in Claude Code.
[ ] Trust the workspace, so it can read, create and move files there.
[ ] Ask in plain language for the scaffolding: what the business is, that context is coming, and that it acts as your marketing team across ads, site and social.
[ ] Paste the three paragraphs first, then product facts, the angles you already use, and the way you talk about the thing.
[ ] Confirm CLAUDE.md reads like an operating document: what this project is, and which files a future session reads first.
[ ] Confirm the context set exists as separate files: brand positioning, the customer avatar, product facts, voice and copy, a message bank, playbooks, and the competitive picture.
[ ] Confirm there is an assets folder for what you supply and an output folder for what it produces.
[ ] If you stall on what to paste, run an interview-style skill: it questions you one topic at a time and writes the answers into a project file.

Treat CLAUDE.md as living: when you learn something that changes how the marketing should work, have the change written in.

2. Brand material, then the count

A file that appears in the sidebar is not a file the model has read, and the failure surfaces an hour later as generic imagery.

[ ] Drag in the logo in light and dark, your product photography, and the brand guidelines PDF.
[ ] Fix the number in your head first: you know how many product shots went into that folder.
[ ] Ask it to inventory the folder and list back what it can see, then confirm the count matches.
[ ] Re-run the count after anything moves. A shifted file is the usual reason an asset drops out of a batch.
[ ] No logo yet? Ask for a plain wordmark in the colors you have settled on, and write the palette and typefaces into a one-page guidelines file.
[ ] No product shots yet? Generate a set from a written description, keep it in a separate folder, and never let it ship as photography of a real product.

Ten seconds of counting decides whether the next hour runs on your assets or on the model's imagination.

3. The media generator, connected over MCP

MCP is a standard socket that lets an assistant operate another service, so you brief the generator from inside the project.

[ ] In the generator, find the MCP or CLI section and copy the connector URL meant for Claude.
[ ] In Claude, open the settings, go to Connectors, and add a custom connector.
[ ] Paste the URL, name the connector something you will recognize later, and add it.
[ ] Sign in when the prompt appears, then grant the two services permission to talk.
[ ] Verify with one throwaway image and confirm the file is in the output folder. A chat message saying it was made is not a file on disk.
[ ] If the file is missing, fix the connection before briefing anything larger.

4. The first job: the website

I build the site first because it exercises everything at once. Guidelines, persona, assets and copy all have to work together for the page to come out right.

[ ] Find a public repository that builds the kind of site you want, and copy its address in as the technical base.
[ ] Brief the job: the business, the assets you already supplied, and an instruction to read the three paragraphs and the persona before writing copy.
[ ] Check that the design system came out of the guidelines document instead of being invented. Colors and logo treatment should look like your brand, not a template.
[ ] Check that the page opens on the pain and the product arrives a few scrolls down. Opening on the product means the context was not read.
[ ] Expect blank sections wherever you supplied no facts. The model will not invent facts, so the fix is to supply them.
[ ] Two limits before you show it to anyone: it runs on localhost, so that address opens for nobody else until you host it, and a first pass is a first pass.

5. The four formats, and the prompt shape

You do not run these one at a time. Fire several briefs into the same conversation and let them work in parallel.

[ ] Static ad creatives for one promotion: the product render consistent across every frame, brand colors and typography applied, several left without copy on purpose.
[ ] A music-driven sizzle reel out of the generator's marketing studio, in both vertical and landscape cuts.
[ ] Instagram carousels of seven to eight slides, each on one angle: why the product exists, who it is for, what it removes, the flavor line-up.
[ ] Vertical UGC-style video, staged: characters as stills, then storyboards from those characters, then video from the storyboards, then a quality pass that rejects the bad takes.
[ ] Characters before storyboards, storyboards before footage: that order is what stops a clip drifting between shots.

One prompt shape covers all four formats. Copy this block and fill it in.

Motivation. What is happening and why you need this. "There's a promotion coming up, here is the offer."
End goal. What you want back, in countable terms. "A spread of static creatives to test, using the brand context."
Filing. Where the output goes. "Save them in the output folder under this campaign name."
Cost. "Tell me what this run cost."

[ ] Brief in plain language, the way you would brief a colleague who already knows the business, because by now the project does.
[ ] Write the ratios down after a few rounds: stills are the cheap end, carousels sit above them, a music-driven video costs several times more than a still, and staged video is the most expensive because it pays for images, storyboards, footage and a review pass.
[ ] Use the ratios to decide which format you run at volume and which you run once per campaign.

6. Turn the hits into skills

Saving what worked is the habit that makes the setup accumulate instead of resetting every month.

[ ] After each round, name what worked and ask for a skill of it: "I like this carousel style, make a skill so I can produce that look again."
[ ] Name what missed just as explicitly, and rule it out: "never produce them like that again."
[ ] Confirm the skill file landed, then run the next job by calling it by name instead of re-describing the look.
[ ] Keep the character cast in its own folder. The same faces can front the next campaign.
[ ] Keep the creatives you left without copy. The next promotion swaps a headline instead of regenerating the batch.

7. The tracker, and the weekly loop

Volume buys you attempts. Measurement tells you which one to repeat. The tracker connects them, and it does nothing until real numbers go in.

[ ] Ask for a spreadsheet and describe the tabs you want instead of accepting a default shape.
[ ] Generation log: date, type of run, campaign, channel, primary angle, deliverables, model, cost, status.
[ ] Creative list: every asset with the copy that ran on it.
[ ] Angle performance: one row per angle, so you compare angles rather than individual images.
[ ] Copy bank: the lines worth reusing, pulled out of what you shipped.
[ ] Cost summary: per run and per format.
[ ] Feed in real performance figures as they arrive, views, engagement, conversions. Until then the sheet is a log, not a decision tool.
[ ] Hand the sheet back once a week and ask what worked, what did not, and for more of the first kind.

Before you call it done

[ ] The three paragraphs are in a project file, and a cold reader could tell your buyer from a generic one.
[ ] You asked for the asset inventory and the count came back matching.
[ ] One test image from the connected generator is sitting in the output folder.
[ ] The site's colors and typography trace to the guidelines document, and the first screen names the pain.
[ ] One result you liked is saved as a skill you can call by name.
[ ] The tracker holds at least one row where a real performance figure sits beside the generation record.