How to Setup Your DTC Brand's AI Workflows

Over the last month, I’ve had calls with 6 high-caliber agencies and 7-8 figure ecommerce brands. Every single one was asking me some version of the same question:

Or

So I started asking what their setups looked like. And I kept seeing the same pattern.

One mega prompt. Or one mega context file. Or one mega skill that tries to do everything: analyze performance, draft emails, plan campaigns, write ad copy, review flows. All stuffed into one massive document that gets pasted into every conversation.

It's the AI equivalent of hiring one person and telling them "you're the marketing department now." It kind of works. Sort of how the Echo Smart Glasses kind of worked.

So I walked them through the same thing I’m going to walk you through today: a simple way to setup AI agent systems that will work well for marketing in ecommerce., and how to set yours up so the output goes from "meh" to "actually useful."

To be clear, this isn’t how I am setting up my eCommerce brand (I’ll cover that in a future email), but it is a simple and very effective start.

Before we get into the structure, let's talk about where to build this.

Cowork (if you're nontechnical): Anthropic launched this earlier this year and it's genuinely good for ecommerce operators. It runs on your desktop, reads and writes to folders you specify, and can handle multi-step tasks without you babysitting it. No terminal. No code. You just tell it what to do in plain english. It also has plugins now, so you can bundle your skills and connectors into one install.

Claude Code (if you're more technical): Same underlying agent architecture, but you're in the command line. More control, more flexibility. If you're comfortable in a terminal, this is the move. Routines are an amazing unlock here.

The web client (Claude.ai, ChatGPT, etc.): If you're not ready for either of the above, you can still apply everything I'm about to cover. You'll just be doing some manual pasting. It works, but there are limitations. You'll lose the automatic skill selection and sub-agent coordination that makes the other two powerful.

If you noticed a trend with the above… I’m talking about Claude. ChatGPT recently rolled out skills which you can also use. But I use Claude pretty exclusively at this point.

Now let's get into the actual setup.

Most brands I talked to had a single "brand bible" they paste into AI. 3,000 words covering everything from brand voice to customer personas to product specs to competitor positioning.

The problem here is when you give AI everything at once, it doesn't know what to prioritize. Ask it to write an email subject line and it's thinking about your competitive positioning. Ask it to analyze your flow performance and it's stuck on your brand voice. Context overload kills output quality the same way it kills human productivity.

Instead, break your context into layers:

Layer 1: General Brand Context This is the stuff that almost every task needs some awareness of. Your brand voice, your target customer personas, your positioning. Keep this tight. I'm talking 500-800 words max, not 3,000. If your brand voice doc is longer than one page, it's too long for AI to use effectively.

How to build it: Take your best performing emails, your top ads, your product pages that convert. Paste them into Claude and say:

That file becomes your always-on context. In Cowork, you drop it in your project folder. In Claude Code, it goes in your /context/ directory. On the web, you paste it at the top of relevant chats.

Layer 2: Task-Specific Context This is the context that only matters for specific jobs. Your last 90 days of email performance data. Your current ad creatives and their ROAS. Your flow analytics. Your competitor's recent campaigns.

You don't want this polluting a copywriting task. But you absolutely want it available when your AI is doing performance analysis.

In Cowork and Claude Code, these live as separate files that skills can pull in when they're relevant. On the web, you paste them in when the task calls for it. This is called progressive disclosure and it preserves the context window, improving output quality. If none of that last sentence made any sense, just slot it into the “trust Nathan” category.

The key insight here: your AI should get the minimum context it needs for the job at hand, not everything you know about your brand. Think of it like briefing an employee. You don't hand your email copywriter a spreadsheet of last quarter's ROAS data. You give them the brand voice guide and the campaign brief.

A massive unlock here too is really going deep on your ICP context. Don’t go high level, instead go hyper detailed. How do they talk? What words do they use? Common pains they react to? Do they usually respond to promotions? If you want to accelerate this, hitup reddit for common language they might use.

If you read nothing else from this email, read this.

This is the part that changed everything for the brands I worked with.

Instead of one mega skill that tries to handle all your marketing tasks, break your skills into individual job functions. Think of each skill like hiring a specialist.

Here's a starting set I'd recommend for most ecommerce brands:

Email Performance Analysis Skill: Analyzes your Mailchimp (for real though, it’s getting good for eCommerce now)flow, campaign, and SMS data. Looks at open rates, click rates, revenue attribution, subscriber drop-off points. Outputs specific findings and recommendations. This skill pulls in your brand context AND your email performance data.

Email Drafting Skill: Writes email copy for campaigns or flows. This one pulls in your brand voice context and your customer persona data, but NOT your performance analytics. Pro move is to have it with a little performance data around top performing campaigns but not the full set. Its job is to write, not analyze.

Ad Angle Generation Skill: Takes in your product info, customer reviews, competitor positioning, and generates ad angles with hooks and copy variations. You could break this down further by channel (Meta, Google, TikTok) since each channel has different creative requirements. If you haven’t seen Meta’s new MCP yet. You should apply that here as well.

Campaign Planning Skill: Higher level strategic thinking. Takes in your performance data across channels, your product calendar, your goals, and maps out a campaign plan. Pulls in broad context. Comes up with the overall campaign plan.

Performance Analysis (Ads) Skill: Separate from email analysis. Analyzes your ad account data, identifies what's working, what's fatiguing, and where to shift spend. Could be one skill or broken out by channel.

SMS Drafting Skill: Similar structure to email drafting but with the constraints and best practices specific to SMS (character limits, tone, timing considerations).

Each of these skills has two components:

Here's where it gets good.

In Cowork or Claude Code, once you've set up your context files and skills, you don't need to manually orchestrate anything. You just tell Claude what you want:

Claude reads your request, determines it needs the Email Performance Analysis skill AND the Email Drafting skill, pulls in the right context for each step, kicks off the analysis, reads the results, and then drafts copy informed by what it found. Two skills, right context, one request from you.

Claude pulls in the Ad Angle Generation skill, grabs your ad performance context and customer review data, and gets to work. It doesn't waste tokens processing your email flow analytics because that skill doesn't call for it.

If you want a specific recurring playbook (say, every Monday you want a full analysis > ICP review > campaign planning > drafting pipeline), you can create an orchestration skill that calls your other skills in sequence. Think of it as a skill of skills. Slam it into the new routine support and you have your Monday morning marketing briefing, automated.

Three reasons:

Better output quality. When AI focuses on one job with the right context, the results are dramatically better than when it's juggling everything. The brands I talked to who broke their setup out like this saw immediate improvements. Not incremental. Noticeable.

Flexibility. Need to add a TikTok ad skill? Create one new skill file. It doesn't break anything else. Need to update your brand voice? Update one context file and every skill that references it gets the update. With a mega prompt, every change risks breaking something.

Advanced capability. When your skills are clean primitives (individual job functions), you unlock the ability to chain them, run them in parallel, schedule them, and build more complex workflows. That's where this gets really powerful, and we'll dig into that in a future edition.

Don't try to build all of this at once. Start here:

The brands I’ve talked to that are moving to this didn't build it all in a weekend. They built it one skill at a time, improving each one as they used it. That's the move. Otherwise you’re just going to feel overwhelmed.

This Week's Rabbit Holes

And that's it for this week's edition. Stop asking one AI to be your entire marketing department. Break it into specialists, give each one only the context it needs, and watch the output go from "I need to rewrite all of this" to "I need to tweak a few things." One skill at a time. Start this weekend.