Hey there 👋,
We keep seeing the same pattern.
Operators spend 20 minutes re-explaining newsletter formatting requirements to Claude. Same structure. Same tone guidelines. Same output format.
Every single time.
They're treating AI like a blank slate instead of a trained assistant.
The difference: workflows vs. prompts.
The Context Window Problem
Here's what happens when operators don't build workflows.
They write long custom instructions. They embed knowledge bases into projects. They paste the same examples into every conversation. Then they wonder why Claude forgets things or why their responses degrade after a few exchanges.
They're burning context on repetition instead of reserving it for actual work.
Between custom instructions, projects, and MCP servers, most operators are choosing the wrong tool for the job. Custom instructions eat context on every conversation whether you need them or not. Projects lock you into static workflows. MCP servers can consume hundreds of thousands of tokens connecting to external APIs.
The solution: portable workflows you invoke only when needed.
Claude Skills: The Middle Ground
Claude added a feature called Skills. Think of them as flash drives for AI workflows.
Upload a markdown file with specific instructions. Give it a name. Now you can invoke that expertise in any conversation without consuming context until you actually use it.
The setup takes under a minute. Create a skill.md file with your workflow instructions. Zip it. Upload to Claude. Toggle it on.
Now when you type "use the newsletter formatter skill" in any chat, Claude loads those instructions and executes. When you don't mention it, it doesn't consume any context.
This solves the problem where operators either:
Hardcode everything into projects (too rigid)
Recreate prompts every time (too slow)
Load custom instructions that eat context whether needed or not (too wasteful)
Skills give you reusable expertise that's portable across conversations.
Skills vs. Projects vs. MCP Servers
Custom instructions and projects consume context whether you use them or not. If you have 200K token limit and your project instructions use 50K, you're left with 150K for actual conversation. Every time.
Skills cost zero context until you invoke them. You can load 10 different skills and only pay the context cost for the ones you actually use in a given conversation.
MCP servers connect to external data sources (GitHub, databases, APIs). Powerful but can burn through context fast. The GitHub MCP server can consume hundreds of thousands of tokens in minutes just pulling repository data.
When to use each:
MCP servers: Connecting to external APIs or databases where you need live data.
Projects: Instructions that apply to every single conversation regardless of task. Your core operating principles.
Skills: Reusable workflows for specific tasks that you don't need in every conversation. Creating reports, formatting content, analyzing data, generating specific document types.
The breakthrough: you can use skills inside projects. Your project sets the foundation, skills add specialized capabilities you invoke as needed.
The Automation Pattern
Here's how operators are actually using this:
Newsletter formatting: Instead of pasting "use this tone, this structure, these section headers" into every draft, they created a newsletter-formatter skill. Now they write raw content, invoke the skill, get formatted output. Saves 15 minutes per issue.
The operators stuck doing manual work are solving for the wrong constraint. They're optimizing individual prompts when they should be building reusable systems.
The Efficiency Multiplier
Operators who build workflows operate differently than operators who write prompts.
They document processes once instead of recreating them constantly. They invoke expertise as needed instead of loading everything into every conversation. They preserve context for actual thinking instead of burning it on repetition.
They build leverage instead of working harder.
Your newsletter business isn't measured by how many prompts you write. It's measured by how much value you create per hour spent.
Workflows multiply that value. Prompts don't.
🧠 OPERATOR INTEL
8 tasks to finish before January hits
Dan Oshinsky's end-of-year checklist for newsletter operators: audit your automations (broken links and outdated welcome emails are costing you), set up Google Postmaster (free deliverability insights most operators ignore), run a reactivation series (clean lists save money and keep you out of spam), and figure out your big bets now. His Copenhagen workshop sold out in weeks because he announced in January when budgets are fresh.
Metro built 300K newsletter subscribers using print as a funnel
The free commuter paper runs 17 newsletters serving 2.5M daily print readers. Their growth engine: personalized fact boxes that feel like a journalist talking directly to you, plus WhatsApp channels growing 12.5% monthly with 50% of reach from shares. Their AI policy: ideation and subject lines only, never content. "Editorial control and personal voice" is the line they won't cross.
How to use AI agents for marketing
SafetyCulture handles 500,000 free signups annually from 180 countries. Their AI BDR workflow: fetch lead name and company from Salesforce, pull page views from HubSpot to understand intent, check employment history via ZoomInfo to see if they used SafetyCulture before, select two relevant customer examples based on industry and country, then compile the email and add to Gong Engage flow. The system writes personalized outreach at scale without manual research. Results: 3x meeting booking rate, 2x opportunities created. The constraint: costs creep when every reply triggers AI research, so they prioritize high-fit customers.
From tasks to systems: A practical playbook for operationalizing AI
Rachel Woods built a framework for process automation: Clear Picture (document the workflow), Realistic Design (carve out minimum viable portion), AI-ify (build the automation), Feedback (iterate based on test runs), Team rollout (train and maintain). Key principle: automate in small chunks instead of trying to AI-ify an entire process at once. Start with "tiny but useful" - get one piece working well before expanding scope. The approach treats AI like training an intern: give clear instructions for each step, test execution, provide specific feedback, then move to the next step.
AI is eating marketing. Don’t get eaten with it.
Join thousands of CMOs and founders reading AI Ready CMO — the daily memo on how to use AI to scale growth, automate ops, and make better decisions.
Free AI marketing tools, daily updates, and a community to 100x marketing output with AI.
What topics would you like us to explore next?
Reply and let us know:
Content workflows (how to build copilot systems)
Technical deep dives (context engineering, tool design)
Research process (how we validate insights)
Business models (operator economics).
– Richard & Maciej