Turn Claude Into a One-Person Marketing Team (Nate Herk)

Source: raw/Turn_Claude_Into_a_One_Person_Marketing_Team_in_38_Mins.md — Nate Herk YouTube walkthrough, ~38 min (https://www.youtube.com/watch?v=yCACmFTiCto)

A start-from-zero build of a marketing function inside Claude Code Desktop, aimed explicitly at non-technical operators. The demo brand is fictional (Perkform, a canned protein-coffee), and every asset in it — logos, packaging, product shots, website, ads, carousels, UGC video — was generated by Claude driving Higgsfield through an MCP connector. The value is not the generations; it is the project scaffold and the context discipline that make the generations on-brand, plus the first set of dollar-denominated cost figures the wiki has for this workflow.

Key Takeaways

• The scaffold is the product. One folder, opened in Claude Code, containing CLAUDE.md (operating doc), context/ (brand positioning, avatar, product facts, voice and copy, message bank, playbooks, landscape, open questions), assets/ (logos, product shots, brand-guidelines PDF), and output/ (deliverables, one sub-folder per campaign). Every later prompt is short because the context is already on disk.
• The Three Ps are the minimum viable brand context: Pain, Person, Promise. What pain does the business solve, who specifically feels it, how does the brand promise to solve it. The source’s claim is that this is what stops output reading as generic AI slop — assets need “a purpose and intent behind them.”
• “Claude Code” is a misnomer for this audience. The reason to use it over the chat surface is that it works out of a directory, so the project accumulates context across sessions. No coding happens in the demo.
• A brand-guidelines PDF in assets/ propagates further than expected. Claude pulled the design system “straight from your guidelines PDF rather than inventing one” — and then applied the same palette to an Excel tracker it built later, unprompted. The creator notes Claude does not normally style spreadsheets that way.
• Costs, in dollars, for one campaign round: 18 static ad creatives = $3.43\ast \ast ;fiveInstagramcarouselsof7–8slideseach=\ast \ast$9.56; one sizzle reel (vertical + landscape) = ~$17.55. UGC video was the most expensive line because it generated characters, storyboards, videos, and a QA pass. Video is where the money goes; stills are near-free by comparison.
• The UGC pipeline self-QAs and rejects its own output. Order: character images → storyboards → video generation → a QA pass that screenshots frames and rejects failures before reporting done. One generation was rejected over a back-panel/label defect. This is the “verify the artifact, don’t trust the report” discipline arriving inside a creative pipeline.
• Ask for the cost in the prompt. Every task prompt in the demo ends with “tell me how much this costed” and “save it in the correct folder.” Cost accounting and file placement are prompt-level habits, not afterthoughts.
• The closing artifact is a spreadsheet, not an asset. A creative tracker with tabs for generation log, creative assets, angle performance, copy bank, and cost summary — designed so performance data can be fed back weekly and the winners regenerated. This is what turns a generation session into a loop.
• Honest limitation, stated by the creator: the UGC videos “are still very salesy… you can tell it’s not completely organic.” The fix he proposes is to have the agent analyse existing organic UGC and write human-approved scripts before generation, rather than prompting for “organic feel.”

The build, in order

1. Assemble the Three Ps before opening Claude

For the demo brand: the person is a busy fitness-minded professional who relies on coffee, cares about protein, and rushes breakfast; the pain is that mornings require too many separate products (coffee, protein shake, something portable); the promise is coffee and protein in one grab-and-go can.

2. Open an empty folder and let Claude write its own scaffold

The kickoff prompt names the business, the goal (“you’re going to be like my marketing team”), and asks Claude to “create me a CLAUDE.md file as well as any other context files that you think are relevant that you can reference later” — then pastes in the Three Ps and any other business context. Claude produces the CLAUDE.md, the context/ files, and empty assets/ and output/ folders.

The resulting CLAUDE.md is short and does one job: name the business, name the owner, and say “Read these files before writing anything.” It is a router into context/, not a container of context. Treat it as a living file — “Hey Claude, by the way, add that to the CLAUDE.md.”

3. Use an interview skill when you cannot articulate the context

The creator’s “Grill Me” skill is a markdown file whose stated job is to “relentlessly interview the user about every aspect of a topic until you reach shared understanding” — the agent keeps asking one question at a time and saves the interview as context. Install it, run it against your product and marketing, and the output becomes the context/ files you could not write yourself. Distributed through the creator’s own community rather than a public repo, and no source is credited in the video.

Use the canonical version instead. This wiki has documented Matt Pocock’s `grill-me` since May 2026 — MIT-licensed, with the verbatim prompt captured, plus a continuous-checkpoint variant that writes after every answer into a brainstorms/ folder so an hour-long interview survives context overflow. The descriptions match closely enough that these are probably the same skill.^[inferred — the video credits no source] Take the checkpointed version. Background on why this shape crossed from engineering into marketing: Context-Extraction Skills.

4. Drop real assets in and verify Claude can see them

Logos, product shots, and the brand-guidelines PDF go into assets/. The demo hits a real gotcha here: a dragged-in sub-folder rendered oddly in the file sidebar. The creator’s response is the generalisable move — ask Claude what it can actually see (“there should be 11 of them in there”) rather than trusting the UI. Claude read all 11 PNGs and inventoried them; the sidebar was the thing that was wrong.

5. Connect Higgsfield as an MCP connector

Higgsfield → MCP and CLI → copy the Claude connector URL → in Claude, Customize → Connectors → Add custom connector → paste, name it, add, then Connect and authorise. After that Claude drives Higgsfield’s models and Marketing Studio directly; the creator’s framing is that you never navigate the Higgsfield UI again. Setup matches the connector flow already documented in Higgsfield MCP.

6. Fire several campaign prompts in parallel

The demo runs four jobs in one session — BOGO ad creatives, a Marketing Studio sizzle reel, Instagram carousels, and UGC video ads. Each prompt follows the same shape:

motivation → end goal → which tool/model to use → where to save it → “and tell me how much this costed.”

The website step is separate and worth noting: the creator pastes a public GitHub repo URL (Scroll World, a scroll-driven 3D site template by a creator credited as “Oso”) into the prompt alongside the brand context, and asks Claude to combine the repo, the brand guidelines, the Three Ps, and the personas into a site. The result opens on the pain rather than the product, and uses Higgsfield-generated Seedance 2.5 clips as scroll-linked hero video. It runs on localhost — publishing is a separate GitHub → Vercel step the video defers.

Costs and what they buy

Deliverable	Volume	Cost
Static ad creatives	18 images (incl. blank/no-copy variants)	$3.43
Instagram carousels	5 carousels × 7–8 slides	$9.56
Sizzle reel (Marketing Studio)	1 vertical + 1 landscape	~$17.55
UGC ads	4 characters + storyboards + videos + QA pass	highest line item; not itemised in the source

Models named across the run: GPT Image 2 and Nano Banana Pro for stills, Seedance 2.5 for video, Soul for some generations, plus Higgsfield Marketing Studio for the music-and-cuts sizzle reel. The creator’s honest note on the vendor: Higgsfield “is not the cheapest tool out there for all of these different models, but it has everything aggregated into one spot.”

Why the outputs are reusable

The blank-copy variants matter more than they look. With 18 branded plates saved, the next promotion is a copy swap, not a regeneration — and the reverse works too (“I really like this copy, put it on all 18 of these graphics”). The creator’s framing: volume negates luck, and once assets are in market you have the data to decide which angle to make more of.

That is what the creative tracker is for. Its tabs — generation log (date, type, campaign, channel, primary angle, angles covered, deliverables, generations billed, tool/model, credits, dollars, status), creative assets, angle performance, copy bank, cost summary — are designed so a weekly agent pass can add view/engagement/conversion numbers and then generate more of what worked.

Try It

1. Write the Three Ps down first, in one paragraph each. If you cannot, run an interview skill against yourself before you generate anything. Everything downstream inherits this.
2. Create the folder, then let Claude build the scaffold rather than hand-authoring CLAUDE.md. Ask for CLAUDE.md plus “any other context files you think are relevant.”
3. Put a real brand-guidelines PDF in assets/. It is the highest-leverage single file in the folder — the source shows it propagating into web design, ad plates, and even spreadsheet styling.
4. Add “and tell me how much this costed” to every generation prompt from day one, and name the output sub-folder. Retrofitting cost tracking after 200 generations is the expensive way to learn this.
5. Start with stills, not video. At ~$3.43for18platesversus$17.55 for one sizzle reel, the static-ad loop is where you validate the context scaffold cheaply.
6. Ask Claude to verify what it can see after any bulk asset drop, before you spend credits generating against those assets.
7. Do not accept the first UGC pass as organic. Have the agent analyse real organic UGC in your category and write scripts you approve before it generates video.
8. Turn a run you liked into a skill. The creator’s closing move — “every time you create an ad like this, use this skill” — is what makes round two cheaper than round one.

Related

• Higgsfield MCP + Claude Code — End-to-End Ad Agency Workflow — the engineering-flavoured sibling of this build (Firecrawl brand brief, Claude-as-critic scoring, credit-denominated costs); read both, they cover different halves
• Higgsfield MCP — the connector, its model roster, and the SDK alternative
• Higgsfield → Figma — Editable AI Static Ads — the fix for on-image typography, which this workflow does not solve
• Higgsfield Virality Predictor — scores short-form output; the natural gate before the tracker’s “angle performance” tab
• Context-Extraction Skills — how the interview-skill pattern reached marketing from an engineering skills repo
• Matt Pocock’s Skills Repo — the canonical grill-me, with the verbatim prompt and the checkpoint variant
• CLAUDE.md Primer — why the router-not-container shape used here is the right one
• How Alex Hormozi Uses AI — The 4-Pillar System — same creator, the strategy layer above this execution layer
• Agent Skills Overview — turning a good run into a reusable skill
• AI Ad Creative Testing & Optimization — where the creative tracker’s performance loop should end up
• AI Marketing — topic index

Open Questions

• Which model produced which asset. The tracker has a tool/model column and the creator names GPT Image 2, Nano Banana Pro, Soul, and Seedance 2.5 across the run, but the video never attributes a specific model to a specific deliverable. Per-model cost/quality for this exact job is unresolved.
• UGC cost. Called “the most expensive” but never given a number, unlike the other three deliverables.
• Scroll World provenance. Credited on-screen to a creator named as “Oso” with a public GitHub repo, but the auto-caption gives no verifiable handle or repo path. Unverified — do not cite the repo without checking it.
• Does the scaffold survive multi-SKU brands? The demo is one product line with three flavours. The sibling article raises the same question; neither source answers it.
• The Excel-styling behaviour. Claude applying brand colours to a spreadsheet without being asked is reported as surprising by the creator. Whether that is the brand-guidelines PDF being read as a general style source, or a one-off, is untested.