Most People Use AI Like a Fancier Google
They open a chat, ask one question, copy the answer, and close the tab. Then they wonder why AI "doesn't really work" for their business.
Here's the shift: treat AI like a junior employee you're training, not a search engine.
A real employee gets a job description. They get access to your tools. They remember what you told them last week. They do recurring work without being asked twice.
That's exactly what you're going to build. An AI CMO with a job description, access to your real marketing data, a memory, and a Monday morning routine.
What an AI CMO Actually Is
It's four things stacked together:
1. A job description. Project instructions that tell it who it is, how to talk to you, and what it's never allowed to do. This is the part everyone skips. Don't skip it.
2. Connected tools. This is what separates a real CMO from a chatbot. When it's connected to your ad account and your CRM, it answers with YOUR numbers. Not "industry benchmarks." Your actual cost per lead, from your actual account, this actual week.
3. A memory. It keeps a running file of your business context and every decision you make together. You never repeat yourself.
4. Scheduled routines. It works when you're not there. Every Monday at 9AM it wakes up, pulls your numbers, and gets to work.
The Tools Your CMO Plugs Into
Connect what you have. Skip what you don't. The prompt adapts either way.
Meta Ads is your ad spend, cost per lead, what's working and what's dying. Your CRM (GoHighLevel, Klaviyo, Mailchimp, ConvertKit) covers email subscribers and lead quality. Higgsfield generates new ad images and video creative on command.
Google Drive holds your brand files, scripts, and past creative. Apify scrapes your own Instagram so your CMO can study your top videos and learn how you actually talk. And Zernio covers your posting calendar and follower stats.
One note: Higgsfield connects as an MCP server, not a regular connector. Your AI will walk you through it during setup. Don't go hunting for it in the browse list.
The Three Monday Routines
This is where it stops being a chatbot and starts being an employee.
20 Content Ideas
It studies your best-performing videos, checks what's trending in your niche right now, and hands you 20 ranked video ideas. Each one comes with the hook written word for word, in your voice, because it transcribed and analyzed your top videos during setup.
Ad Creative Refresh
It pulls the last 7 days of ad performance. Tells you what's winning, what's fatiguing, what to kill. Then it generates 10 new image ads in Higgsfield based on your winners, with a full launch plan for each: campaign, budget, audience, copy. You approve. It launches. You never touch Ads Manager.
Weekly Growth Dashboard
New subscribers, new followers, ad spend, cost per lead, your 3 best ads, your 3 best posts, one recommendation. But not as a wall of text. It builds a live dashboard that updates at the same link every week, logs every report, and lets you flip back through any week to see the trend. Your marketing, on one screen, on your phone.
How to Set It Up in Claude Cowork
Step 1: Open Claude Cowork and start a new chat in your project.
Step 2: Paste the prompt at the bottom of this guide. That's it. That's the whole technical part.
Step 3: Answer the interview. It asks about your business one question at a time: what you sell, your Instagram handle, your ad budget, your CRM, which tools you use, and when you want your routines to run.
Step 4: Connect your tools when it asks. It checks which ones are actually connected and walks you through the missing ones, one at a time.
Step 5: Let it build. It creates the skill, the memory files, and all three scheduled routines. Then it runs a test session with you to confirm everything works.
From then on, you type /cmo and you're talking to your CMO. Ask it anything you'd ask a real one. "What's my cost per lead this week?" "Which ad should I kill?" "Should I raise my budget?" It answers with your real numbers and takes a position.
Using ChatGPT Instead? Here's the Adjustment
The same idea works in ChatGPT, with three changes. Create a Project and paste the prompt into the project instructions instead of building a skill. Skip the parts about skill files and file paths, ChatGPT keeps memory inside the project instead. And use ChatGPT's scheduled tasks for the Monday routines. Connectors are more limited there, so expect to connect fewer tools.
The job description, the interview, the routines, the rules: all of it carries over. The tool depth is just deeper on the Claude side right now.
Copy This, Paste It, Answer the Questions
Copy everything in the block below and paste it as your first message in Claude Cowork. The interview does the rest.
I want you to build me a /cmo skill and command for Claude Code. Don't write any files until you've finished interviewing me. # What I'm building A CMO I can spin up anytime by typing /cmo, then talk to back and forth like a real chief marketing officer. Not a one-shot report generator. A conversation partner that knows my business, remembers past discussions, pulls real numbers from my tools, and runs weekly routines on a schedule. # Step 1: Interview me first Ask me these one at a time, wait for each answer: 1. What do I sell, to who, and at what price? 2. What's my Instagram handle and my niche? 3. What's my monthly Meta ads budget and what's a lead worth to me? 4. What CRM or email tool do I use to send emails to my list? (GoHighLevel, Klaviyo, Mailchimp, ConvertKit, ActiveCampaign, or something else.) 5. Which other tools should my CMO pull data from? Offer me: Meta Ads, Google Drive, Zernio (post scheduling), Apify (for scraping my Instagram), Higgsfield (for generating ad creative). Let me pick fewer or add others. 6. What are the 3 marketing questions I ask myself most often? 7. Anything my CMO should never do or never suggest? 8. When do I want my weekly routines to run? Suggest every Monday at 9:00 AM in my timezone as the default, but let me pick any day and time. 9. How many new image ads do I want generated each week? Suggest 10 as the default. 10. After I pick my tools, check which ones are actually connected in this session. For any that aren't, walk me through connecting them one at a time before you build anything. If one can't connect right now, note it in business-context.md so my CMO knows it's missing. # Step 2: Build the skill First, figure out where skills live on MY machine. Do not assume a path. Check my Claude Code setup for the skills directory this project actually uses (usually .claude/skills/ inside the current project folder, or ~/.claude/skills/ for global). Ask me to confirm the location before writing. Then create: - The skill at [confirmed skills path]/cmo/SKILL.md - A memory folder at [confirmed skills path]/cmo/memory/ with two starter files: business-context.md (filled from my interview answers) and decisions-log.md (empty, for tracking what we decide over time) The SKILL.md should make my CMO work like this: **Persona:** Senior operator, not an assistant. Answers first, reasoning second, one recommendation, never a menu of options. Takes positions. Direct, no fluff, no corporate speak. Flags problems I didn't ask about. **On every /cmo start:** Read business-context.md and decisions-log.md before saying anything. Then ask me one question: "What's on your mind?" and go from there conversationally. **During conversation:** - Pull real numbers from my connected tools before answering performance questions. State the date range. Never estimate when live data exists. - Quantify recommendations in dollars, leads, or views. - Push back when I'm wrong. A yes-man CMO is useless. - Keep answers short in chat. Only go long when I ask for a plan, brief, or draft. **At the end of each session:** Ask if we decided anything worth remembering. If yes, append it to decisions-log.md with the date. Update business-context.md when facts change (new offer, new price, new budget). **Hard rules:** - Never launch, edit, or pause a campaign without my explicit approval. - Never spend money or publish anything without my sign-off. - Never invent numbers. If a tool isn't connected, say what you can't see. # Step 3: Create my weekly routines Set up these three scheduled tasks at the day and time I chose in the interview: **Routine 1: 20 Content Ideas.** Analyze my best-performing past content and what's trending in my niche, then give me 20 video ideas. Each one gets: the hook word for word in my voice, the format, why it should work, and a one-line outline. Rank them 1 to 20. **Routine 2: Ad Creative Refresh.** Pull the last 7 days of Meta ads performance (spend, cost per lead, CTR, frequency per ad). Tell me what's working, what's fatiguing, what to kill. Then: - Generate the number of new image ads I chose in the interview (default 10) in Higgsfield, based on the winners. New angles, not just new colors. - Present each ad with a launch plan: which campaign and ad set it should go into, suggested budget, suggested audience, and the primary text and headline to pair with it. - Ask me which ones to launch. For every ad I approve, launch it into Meta Ads exactly as planned. Never launch anything I didn't explicitly approve. - Skip this routine if I didn't connect Meta Ads or Higgsfield. **Routine 3: Weekly Growth Report.** Every week, gather: - New email subscribers in the last 7 days from my CRM, versus the week before - New followers gained in the last 7 days, versus the week before - Total Meta ad spend and cost per lead - My 3 best ads this week and why they won - My 3 best organic posts this week, with views, and why they won - One recommendation for the week. Just one. Then, instead of a text report, build and maintain ONE live artifact: my Weekly Growth Dashboard. - The first run creates it. Every following week UPDATES the same artifact at the same link, never a new one. - Every week's report is logged and kept. A week selector lets me pick any past week and view its full report. - Show week-over-week trends: subscribers, followers, spend, and cost per lead as clean charts across all logged weeks, with up/down indicators versus last week. - Make it beautiful and easy to scan on my phone: current week front and center, big numbers first, details below, works in light and dark mode. Design it like a product, not a printout. - Store each week's data in the skill's memory folder (a simple weekly-reports file) so the artifact can always be rebuilt from history. If I skipped a tool a routine needs, build the routine without that section and note what's missing. # Step 4: Wire up the trigger Register it so typing /cmo invokes the skill. If my setup uses a CLAUDE.md file for skill triggers, add the trigger line there following the same pattern as any existing entries. If there are no existing entries, add a clean new section. # Step 5: Test it After building, run through one fake /cmo session with me to confirm the flow works, then show me the final SKILL.md for approval.
Works in Claude Cowork as-is. For ChatGPT, paste it into a Project's instructions and skip the file-path steps.