The Architecture of AI-Driven Efficiency: Lessons from the Trenches

In this conversation, Michael Stelzner and Ali Kelly explain that the main barrier to AI adoption is not a lack of tools, but a failure to align AI workflows with existing domain expertise. By mapping their internal processes for news curation, sales page optimization, and content repurposing, they show that the biggest competitive advantages come from using AI to remove operational friction rather than just generating content. Readers who adopt this system-first mindset gain a distinct advantage: the ability to scale output while improving the quality of the customer experience, moving from trying tools to building a digital workforce.

The Hidden Cost of Fast Solutions

Most teams treat AI like a magic wand, throwing a prompt at a model and hoping for an immediate result. Stelzner and Kelly suggest this is a fundamental error. When Kelly tried to automate news curation, she initially relied on standard prompting and scheduled tasks. These obvious fixes failed because they did not account for the specific requirements of her audience.

The breakthrough happened when she treated the curation process as a software development problem instead of a content task. By moving from simple prompts to a custom-coded workflow hosted in GitHub and debugged via AI, she turned a daily, hour-long manual slog into a streamlined, automated process.

I did not know how to do any of this. My coding history was a GeoCities page in high school and with a little bit of HTML you do not need to know how to code, you just need to know how to direct the AI to tell you where to go and what to do.

-- Ali Kelly

Where Immediate Pain Creates Lasting Moats

Stelzner emphasizes a counter-intuitive principle: the effort invested in building an AI-driven system should be significantly higher than the effort required to perform the task manually. He references the Take the Stairs philosophy, noting that while traditional delegation requires 20 times the effort, AI delegation requires only five or six times the effort.

The payoff is not immediate; it is durable. By investing time now to build skills (custom instructions) and projects (long-term context windows), practitioners create a system that compounds in value. This is the 18-month payoff: most competitors will refuse to spend the time building these custom workflows, preferring the quick dopamine hit of a single-shot prompt.

The best way to use AI is to use AI in your domain expertise in the area that you have the most experience in. And the reason why is because you uniquely are able to direct the AI, you are able to tell it what is good and what is bad.

-- Michael Stelzner

How the System Routes Around Your Assumptions

The most significant systems-thinking insight shared is the feedback loop created by their curation tool. When Kelly’s system flags a story as declined, the AI records the reason. Every two weeks, those reasons are used to refine the curation criteria.

This creates a self-correcting system. The AI is not just executing a static set of rules; it is learning the specific, evolving preferences of the audience. By embedding this feedback loop, the process becomes more accurate over time, creating a process that is impossible to replicate with static, one-off prompts.

The Shift from Information to Access

Stelzner notes a critical shift in the market: information has been commoditized. Because AI can synthesize massive amounts of data, the value of a course or training has plummeted. The system has responded by devaluing static information and increasing the premium on access, specifically access to experts, live experiences, and community. Their decision to reposition their sales page was a direct response to this shift, moving away from how to get started to how to build workflows and master AI.

Key Action Items

• Audit your daily repeat processes: Identify the tasks that take up the most time and require high-specificity filtering, such as news curation or data synthesis. Plan to build a custom workflow for these within the next 30 days.
• Stop using static prompts for complex tasks: Transition your most frequent tasks into Projects or Skills (custom instructions) that hold context, style guides, and specific operational constraints.
• Implement a feedback loop: If you are building an AI-driven curation or analysis tool, ensure you have a mechanism to decline outputs and feed those reasons back into the system instructions to refine future performance.
• Leverage your domain expertise: Do not attempt to use AI in areas where you lack experience. Use it to augment your existing craft, whether that is copywriting, coding, or community management, where you have the discerning eye to judge quality.
• Shift your value proposition: If you are selling information, shift your focus toward community and live access. This is a long-term investment that pays off as information becomes increasingly commoditized.
• Debug your code with AI: Even if you are not a developer, use Claude or similar tools to help you clean up code and integrate disparate systems like GitHub and automated fetching tools. This pays off in 12 to 18 months by creating a proprietary, automated infrastructure.