Disclosure: Some links in this article are affiliate links. We may earn a commission at no extra cost to you.

I watched a solopreneur run her entire brand through Claude.ai last month. She opened a new chat for every email, re-pasted her brand guidelines each time, and burned through her Max quota by Wednesday writing the same cold-email variants she wrote the week before. When I asked why she didn't use Projects, she said "I thought that was for coders."

That's the gap. Most solopreneur marketers paying $20 or $200 a month to Anthropic are using maybe 12% of the platform. Claude.ai chat is the surface. Underneath is a stack with Projects, Model Context Protocol (MCP), Skills, the Agent SDK, and Computer Use, and each one removes a different piece of manual drudgery from a one-person marketing operation. The difference between using Claude as a chat toy and running it as an embedded marketing operator is roughly the difference between owning a laptop and knowing how to install software on it.

This playbook walks through the full stack in the order a solopreneur should actually adopt it. No Python required for the first three steps. Tested on my own stack as of April 2026.

What you need before starting

Before touching anything past the chat box, gather these:

• A Claude.ai Pro ($20/month) or Max ($100 or $200/month) subscription. Free tier can't run Projects with files or Skills.
• Brand assets collected into one folder: tone guide, 10 to 20 top-performing emails, your positioning doc, ICP definitions, and any style rules you've already written.
• A Notion workspace if you don't already have one. I push Claude outputs into Notion as the source of truth because Claude Projects don't yet have two-way sync with external databases.
• API access flipped on at console.anthropic.com for steps 4 and 5. This requires a separate billing account from Claude.ai.
• A credit card with about $50 of monthly budget for API experiments. Most solopreneurs will spend under $30.

If you don't have the brand corpus yet, stop here and build it. Claude can't mimic a voice that exists only in your head.

Sonnet 4.6 vs Opus 4.7: the default-picker

Solopreneurs waste more money on model selection than anything else in Anthropic's product line. Opus 4.7 is seductive because it's the newest and smartest. It's also roughly 5x the cost of Sonnet 4.6 per million tokens, per Anthropic's pricing page, and you rarely need its extra horsepower for marketing work.

Here's the per-token math as of this writing:

Model	Input / 1M tok	Output / 1M tok	Use for
Haiku 4.5	$1	$5	Classification, tagging, subject-line A/B, RSS scoring
Sonnet 4.6	$3	$15	Default for 85% of marketing tasks: emails, landing copy, social, repurposing
Opus 4.7	$15	$75	Strategy docs, positioning rewrites, long-form pillar research

My rule: default to Sonnet 4.6. Upgrade to Opus 4.7 only when the output will be used more than 20 times (positioning doc, pillar article, launch narrative) or when the task requires multi-step reasoning across a 100k+ token context. Drop to Haiku 4.5 for any classification task where you're processing more than 500 items a day.

Step 1: Set up your Brand Voice Project

Projects are Claude.ai's persistent-context feature. You upload files, write custom instructions, and every conversation inside that project inherits both. No re-pasting style guides. No "remember we're B2B not B2C" reminders. The UI sits in the left rail of claude.ai and you create one with the + button.

Here's where Evan gets opinionated about the UX. Claude Projects and ChatGPT Custom GPTs look similar on paper, but the file handling is not. Claude Projects retrieve from your uploaded files using full-context injection (up to 200k tokens pulled in per turn, with caching), while Custom GPTs chunk files into a RAG index with a smaller effective context window. For a brand voice task, where you want Claude to read four or five long-form examples in their entirety before writing, Projects produces noticeably more consistent tonal mimicry. I've run blind tests with three clients and Projects wins the voice-match check every time.

The downside: Projects has no scheduled triggers and no public sharing with a custom domain, both of which Custom GPTs do. If you're building a product experience, use GPTs. If you're building an internal voice operator, use Projects.

What to put in your Brand Voice Project

• Custom instructions: ICP, tone rules, banned phrases, and the 2 to 3 things your brand always says differently than competitors.
• Files: 10 to 20 best emails, your About page, your pillar article, your positioning one-pager, your three most common objections with answers.
• A starter turn template: "Write a [asset type] for [ICP segment] about [topic]. Tone should match Email 3 from the corpus. Keep it under [N] words."

Pair this with Notion as your canonical knowledge base. I keep the Notion page versioned, export the markdown weekly, and re-upload to the Claude Project so it stays in sync. This is friction, and Anthropic should fix it. For now, a weekly Sunday reset takes six minutes.

Step 2: Connect MCP servers (beehiiv, Notion, Gmail)

Model Context Protocol is the open standard Anthropic released in late 2024 for connecting language models to external tools. In 2026 it's the plumbing that makes Claude useful outside a chat window. Think of MCP servers as adapters: one for your ESP, one for your CRM, one for your file store. You install them once in Claude Desktop or via a hosted MCP gateway, then Claude can read and write to those systems in any conversation.

The honest comparison: ChatGPT Connectors do the same job, but the MCP spec is genuinely open. According to the Model Context Protocol documentation, over 200 servers exist at the time of writing, and you can write your own in 50 lines of TypeScript. ChatGPT Connectors are a closed list curated by OpenAI. For a solopreneur who wants to tie Claude to a niche tool (a specific ESP, a GitHub repo, a private Postgres), MCP wins on integration speed.

The three MCP servers every solopreneur marketer should install

1. beehiiv MCP server: beehiiv shipped official MCP support in late 2025. Claude can now read subscriber counts, publish drafts, and pull open-rate stats without a Zapier hop. If you're on beehiiv, this is the single highest-ROI install. I wrote a separate walkthrough on the setup at the beehiiv MCP integration guide.
2. Notion MCP server: read and write to any Notion database. My agent pushes weekly content calendars, editorial feedback, and SEO briefs straight into a Notion board. Compared to ConvertKit or a Google Doc workflow, this removes an entire copy-paste step.
3. Gmail MCP server: for draft-only access. Critical distinction. Let Claude draft replies to partnership inquiries but never auto-send. I've caught two hallucinated statistics this way that would have embarrassed me with a podcast host.

Step 3: Build a custom Skill for landing page audits

Skills launched publicly in early 2026. They are packaged instruction bundles Claude loads on demand. Think of a Skill as a saved expert persona plus a set of reference files plus a workflow. Projects set the global context for a workspace. Skills let you spin up specialist modes inside that workspace.

I use one Skill constantly: a Landing Page Audit Skill. It takes a URL or pasted HTML, checks for 14 specific patterns (headline clarity, proof-point placement, CTA specificity, form-field count, loading speed note), and outputs a scored rubric with rewrites. Building it took me 40 minutes.

Building the Skill, concretely

In claude.ai, open Skills from the left rail. Create new. Paste these three things:

1. Name and trigger description: "Landing Page Audit. Use when the user pastes a URL or HTML and asks for a landing page review."
2. Reference files: upload the rubric as a markdown doc. Mine has 14 criteria with pass/warn/fail definitions for each.
3. Workflow instructions: step 1 fetch or accept the HTML, step 2 score each rubric criterion, step 3 output a table with score plus a rewrite suggestion per criterion, step 4 close with top 3 priority fixes.

Now any time I paste a URL in my main Claude.ai window, the Skill auto-triggers. No copy-pasted prompts. No "act as a conversion expert" gymnastics. This is where the stack starts feeling like actual software instead of autocomplete.

For audit work specifically, I pair this with Okara AI when I need mystery shopper traffic on a test page, and Okara's synthetic user reports feed right back into the Skill as an input file.

Step 4: Automate with the Agent SDK

The Agent SDK is the graduation step. You move from running Claude inside claude.ai to running Claude as a headless service inside your own code. TypeScript and Python are both first-class. The SDK handles tool-call loops, permission gates, file-system access, and subagent orchestration out of the box, which means you write business logic instead of plumbing.

A solopreneur-sized use case: a nightly job that pulls yesterday's beehiiv subscribers, checks their email domains against a list of target-account domains, writes personalized LinkedIn opening messages, and saves those messages to Notion for me to review over coffee. That agent is about 120 lines of TypeScript. It runs for $0.30 a night on Sonnet 4.6.

Important honesty: you need to be comfortable with a terminal and reading TypeScript. If that's a no, stop at Step 3 and use n8n or Make to chain Claude API calls instead. Both platforms now support Claude models natively, and n8n's visual workflow builder covers 80% of the Agent SDK's use cases without code. The tradeoff is orchestration ceiling: once you need dynamic subagents spawning other subagents, the SDK is the only place that works cleanly.

Agent SDK vs a no-code workflow builder

I've built the same "repurpose long-form to 10 social posts" pipeline in both n8n and the Agent SDK. The n8n version is 11 nodes and took me 25 minutes. The Agent SDK version is 180 lines and took two hours the first time. But the Agent SDK version handles edge cases (empty input, rate limits, retry on 529, subagent fanout to three platforms in parallel) with maybe 20 extra lines each. In n8n you're stacking conditional nodes forever. If you're running this pipeline 50 times a week, the SDK pays back. Under 20 times a week, stay in no-code.

For a deeper comparison of no-code options, see how AI agents are replacing the marketing stack.

Step 5: Deploy Computer Use for competitor teardowns

Computer Use is Anthropic's feature that lets Claude drive a browser and a desktop directly. It clicks, types, takes screenshots, and reads the screen. When Anthropic first shipped it in late 2024 it was a toy. As of the 4.7 generation it's usable for specific recurring jobs.

My one productive use case: monthly competitor teardowns. I have an agent that logs in to five competitor free tiers, records the onboarding flow screen by screen, writes a narrative breakdown of the UX decisions, and drops it all in a Notion page. This used to be a half day of work every month. Now it runs unattended in about 40 minutes for roughly $4 in API spend.

The technical-execution comparison to OpenAI's Operator matters here. Operator runs in OpenAI's sandboxed browser. Computer Use runs on your machine (or a VM you spin up), which means Claude can touch your actual password manager, your logged-in sessions, and your local files. That's more capable and more dangerous. According to TechCrunch's coverage of the 2025 agent-safety wave, Operator is safer by default; Computer Use is more flexible. For a competitor teardown where I want the agent to use my test accounts, Computer Use wins. For anything touching live customer data or a production tool, I use neither and do it myself.

What Claude still can't do for marketers

Evan rule: be honest about the gaps. As of April 2026:

• No native image generation. Claude can describe images and read them, but it doesn't create them. You'll pair it with Midjourney, Ideogram, or Nano Banana. Don't expect an all-in-one the way GPT-5 handles DALL-E inline.
• No realtime web without MCP. The web-search tool in claude.ai works, but coverage is thinner than ChatGPT's. For news-reactive content, I'd still open Perplexity first.
• Bulk scraping at scale is not its job. Computer Use can scrape ten pages before rate limits and anti-bot systems kick in. For 1,000 pages, use a dedicated scraper like Firecrawl and pipe results into Claude via the API.
• No built-in CRM. It connects to your CRM via MCP but doesn't replace one. If you need outbound sending at volume, look at beehiiv for newsletters and GetResponse for automations.
• Short-form video hooks are not its strength. I use GetHookd for TikTok and Reels hook generation specifically because it's trained on the short-form pattern library Claude wasn't.

Common mistakes to avoid

• Using Opus for everything. You will burn $400 a month and not notice the quality jump on routine copy. Default to Sonnet.
• Starting a fresh chat for every prompt. Projects exist for a reason. Each new chat costs you the context-loading tax.
• Uploading your entire Notion export as one 900-page PDF. Claude retrieves better from 10 focused files than one giant doc.
• Building the Agent SDK before you've mastered Projects and Skills. Each layer teaches you the prompts you'll need for the next.
• Letting Computer Use touch your live accounts. VM or die.

Frequently asked questions

Do I need the Max plan or is Pro enough?

Pro at $20/month covers Steps 1 through 3 cleanly. You hit rate limits on Max tier use around 200 long messages a week. I run on the $100 Max plan and spend roughly 30% of my quota. The $200 tier is only worth it if you're running multiple Projects with heavy file attachments daily.

Can I use Claude to write SEO content that ranks?

Yes, but not by prompting "write an SEO article about X." Feed it your keyword research, competitor SERP analysis, and your brand voice Project. Claude is a better editor than a blank-page writer for SEO. For the full workflow, see AI copywriting for marketing.

Is Claude better than ChatGPT for marketing writing specifically?

For long-form voice mimicry and editing, yes. For fast brainstorm and image generation, no. Most solopreneurs I work with end up paying for both at Pro tier ($40/month total) and routing tasks accordingly. If I had to pick one for pure copy quality in 2026, Claude wins my blind tests 6 of 10.

How do I keep my brand voice from drifting over time?

Version your Project files. Re-upload the latest best-performing email monthly and archive the worst. The voice Project is a living document. I review mine every 30 days and swap out two or three files. More on the repurposing side at content repurposing for solopreneurs.

What's the single biggest mistake I should avoid in the first 30 days?

Skipping the Project setup and jumping to the Agent SDK because it sounds cooler. The SDK is capable and also slow to learn. You'll get 80% of the ROI from a well-built Brand Voice Project plus two MCP servers, and you'll get it in the first weekend.

Tools and resources

• Notion. Canonical knowledge base that pairs with Claude Projects.
• Make. MCP-compatible automation hub for non-coders.
• Zapier. Alternative automation platform with Claude integrations.
• n8n. Self-hosted workflow builder, strong for Agent SDK-adjacent work.
• beehiiv. Newsletter platform with first-party MCP support.
• ConvertKit. Alternative ESP if you're not on beehiiv.
• Okara AI. Synthetic user testing that feeds into Skill audits.

Further reading on the broader stack: the one-person marketing department and AI marketing tools in 2026.

Next steps

Block two hours this weekend. Build your Brand Voice Project in the first hour. Install the Notion and beehiiv MCP servers in the second. That gets you 70% of the value in this playbook before Monday. Everything past Step 3 is compounding returns on the foundation you set today.

For the full curated list of tools that pair with this stack, browse the Ea-Nasir AI tools directory and filter by the integrations you already use.