AI Mistakes We Made (So You Don't Have To)

Full video transcript

This is our AI Profit Accelerator final one of the year, with our summit. And these calls are the members of our community that have been with us for years, tuning in weekly as we discuss new AI updates, insights, wins, and things of that nature that we’re doing as we’re putting AI into practice in the ever evolving world of artificial intelligence.

We’ve got Shelley Berman Rivera, we’ve got Robin, we’ve got Bruce, we’ve got Omar, and we’ve got Kian joining today. As always, this will be our last Scale with AI summit series of the year, and I’m excited to hear the final wins that you guys have to share with us here in today’s summit series. We’ve covered so much over the course of the past year — it’s pretty wild to think about all the advancements that AI has brought, and also what we’ve been able to do with that in our own businesses.

I’d love to open up the floor to you guys. Who wants to kick us off as we do in AIPA fashion, starting off with our big wins of the week, and introduce yourselves so people know who you are, what you do, and a case study you’ve been implementing with AI.

Omar kicked it off: “I know I’ve said it before, I know it sounds like fanboying, but I’m so grateful for you having put together the community from way back when all the way till now. I want to share an AI win — some of you already know this because you’re much more into coding than someone like me who uses AI more in the background for administrative purposes. I was messing around the other day, wondering if it’s possible for me, instead of creating content and then copying and pasting it over into my WordPress site, to directly connect it. In that process I discovered the WordPress MCP, which you can hook directly into Claude Desktop. I was able to literally type a couple of prompts and have it automatically create the content on my WordPress site, save it in draft mode for me to edit. I use Divi as my builder, and I asked if it could move the content into Divi in preparation for further processing — and it did. For me, and probably for a lot of solopreneurs, that’s a game-changer. It is a little bit technical to set up, but not beyond most people in this group. You can download the WordPress MCP from the GitHub for Automattic — just look up WordPress MCP.”

Darby thanked Omar and gave some context: he’s stepping away from running this mastermind group specifically, which is why this is the last session he’ll officially host — though Shelley is carrying the torch and the group will continue to meet. It’s been almost exactly three years since the group opened.

Robin raised a toast: “It’s 7 p.m. in the UK. I’m raising a toast to you and saying thank you for all that you’ve done for this community over the last three years. You’ve singlehandedly wrestled down multiple summits and been extremely consistent in executing and raising these events.” Robin then shared his win — a continuation of a theme he’s mentioned before: hyperspecific newsletters using ChatGPT. He now has eight newsletters automatically scheduled, arriving by email between 7 and 8 a.m. each morning, tuned over a number of days to deliver meaningful market insights, price changes, and product updates. Each newsletter is tied to its own ChatGPT thread; clicking the email link opens the newsletter directly in ChatGPT. He’s now exploring copying these into Substack, taking only five to ten minutes of formatting to serve multiple niche audiences, potentially as paid subscriptions — “a business model in a box.”

Darby then turned to his own update: “So last week I talked about making Claude my [redacted]. This week I’ve been making Sora my [redacted].” He started a fresh Facebook page two days prior, posting AI-generated videos, and one already hit 70,000 views within hours, triggering algorithmic reach. He’s using Higsfield, a third-party tool combining Sora, VO, and other video generators, which currently offers unlimited 8-second Sora videos for paid subscribers without a watermark, and a custom GPT to write and flesh out scripts scene by scene. Separately, for a funnel that teaches people in Asia how to use AI to build apps, he stopped manually modifying GoHighLevel pages and instead finds conversion elements from other funnels he likes — screenshotting them, grabbing the HTML source, and giving both to Claude Desktop to recreate and adapt with his own products and bonuses, then pasting the resulting HTML back into GoHighLevel. That change alone increased how many buyers took his full offer stack; his cost of acquisition is around $5 while average order value is now around $200.

Asked to clarify his process, he explained he does three things: takes a screenshot of the element for the design reference, inspects and copies the specific HTML, and — when the AI doesn’t get the styling right from HTML alone — downloads the entire page’s HTML and CSS as additional source material so Claude can match styles to elements.

Shelley shared next: her involvement has been about a year and a half to two years. Her win was less about a tool and more about the group itself — the high-integrity people she’s met and the client partnerships that have resulted, including work with Kian and communication with Bruce and Robin, calling it “a business model in a box” of trust and collaboration.

Bruce shared several wins. He recalled starting with Jasper years ago, taking one of Darby’s classes, and later learning autoscripting and multi-step frameworks, which he built into a course teaching marketing frameworks to other coaches. His current wins: he wrote a book about AI with AI, aimed at business professionals at small and midsize companies without access to formal AI training, using Ideogram for character images and Napkin AI for framework graphics — with a foreword from Darby. He’s already sketched a follow-up book, The AI Manifesto, exploring twelve beliefs humans need to thrive in the AI era, developed in a single session with Claude. He also runs a marketing agency for health practitioners, using Fathom-recorded interviews to build business documents — a shorter version of Darby’s Market Sauce — covering business philosophy and differentiation, then feeding that into custom GPTs so clients can request content (for example, a script on a specific treatment topic) and receive a YouTube script, blog post, email, and social posts from one input. Lastly, he emphasized the value of a personal knowledge management system (a “second brain,” using tools like Notion, Obsidian, or Evernote), and described using Claude, Claude MCP, and Cursor to organize his own notes and draw connections between them rather than just accumulating unconnected notes.

A member named Grant asked about Claude Skills, prompting Bruce to explain: Claude Skills come preloaded with skills you can import via checkbox, including a “skill creator” skill. In chat, you simply say you want to use skill creator by name — no special commands — and ask it to read a website and build a brand guideline from that content. It scrapes the source material and produces a packaged skill: a folder of markdown documents and examples with an index file, distributable as a .skill file that, when renamed to .zip and expanded, is just a folder of text documents — similar in spirit to a robust custom GPT.

Building on that, Bruce described finding a repo called Skill Seekers through a newsletter, which uses a Python library (Beautiful Soup) to crawl documentation sites, specifically ones with API documentation, and package the result into a Claude-usable skill. He ran it against GoHighLevel’s API documentation to create a “GoHighLevel API” skill, which then gave him step-by-step guidance and even generated dashboard code for that API — despite being a non-developer. He plans to apply the same approach to a client using Company Cam and another using Commerce 7. He noted Skill Seekers also has an MCP so Claude can build the skill via natural language, though its instructions assume a Linux/bash environment, which complicates Windows use.

Bruce credited Kian’s development approach (vibe-coding locally, testing, then pushing to GitHub main connected to a Vercel project for auto-deploy) as the model for his own learning path: rather than learning to code from scratch, he studied known-working starter projects (including OpenAI’s ChatKit starter kit) and used GitHub plus Vercel to get comfortable deploying and modifying real, working demos.

Darby closed the wins portion by emphasizing that putting learning into action, not just accumulating tools and ideas, is what actually moves the needle — and that knowing when to focus and tune out the noise is as valuable as staying open to new tools.

In the open floor discussion, a member (not part of the mastermind) asked whether Darby was leaving Gen AI University altogether or just this mastermind; Darby clarified it’s only this specific weekly mastermind that’s ending in its current form — Gen AI University continues, with changes coming to where and how the community is hosted.

A member named Assad asked whether you could take an expert’s book or course on a topic like Facebook advertising, dump it into a knowledge base, and have an AI agent run that part of the business according to it. Kian answered that yes, AI can be trained on specific resources and can already emulate well-known frameworks (acting as Dan Kennedy or Gary Halbert, for example), but it remains a cumulative aggregation of information without guardrails or structure, so it’s good for generics but not for running a business function as a top-tier specialist. Darby added that the useful path is doing this in steps: have AI extract a marketing framework from the source material, split strategy from tactics, and turn that into a step-by-step process — layering in your own business context, which is the shift from prompt engineering to what’s now called context engineering.

Darby elaborated that AI can lack the nuanced intelligence to execute marketing the way a human would, citing a chatbot he ran for an event that took significant tuning, and a recent AI-voice call styled after Tony Robbins that was impressive but clearly not truly present in conversation. His view: AI can help develop materials, but maintaining control over the actual marketing function is still necessary. Kian added that where AI fails is in judgment; the workaround is training it on specific information while creating very specific rule sets and sub-agents for narrow, well-defined actions rather than expecting broad judgment.

Kian gave a personal example: he took the URL to an expert’s book and website and asked Claude to turn it into an 8-week interactive program with a specific curriculum, audience-specific, producing a deliverable program in about three minutes.

Another member turned the discussion to NotebookLM, describing it as a RAG-based system where you load PDFs, YouTube videos, images, and other text-formatted material into a directory. He gave an example of pulling PDFs of Alex Hormozi’s “Money Models” deliverables (shared by someone on Reddit) into a notebook, and noted NotebookLM’s low temperature/threshold means it won’t hallucinate nearly as much as a standard LLM when answering questions grounded in the sources provided. He described maintaining a 45-page master prompt covering his company’s tech stack, team development, and challenges in a Google Doc, loaded into the same notebook so it prioritizes his own business context before other sources.

Another member said he’s used NotebookLM the same way, treating it like an advisor or coach grounded in a specific knowledge base, and praised its mind-map feature. Another described scraping YouTube videos of books recommended by Gamma founder Grant Lee, loading them into a notebook, asking for a breakdown of key lessons, then asking how to deploy those strategies in his own business, and finally breaking that down into a milestone-by-sprint to-do list. He also uses NotebookLM to hold meeting sources so he can look back and ask what was discussed with a client a month prior.

One member closed by sharing an additional tactic: recording a short video overview after a client call and sending that instead of a written recap, since most people skim or skip emails but will watch a video — useful for initial client meetings and scoping conversations.

The session ended with the group thanking each other for the community, with Darby thanking everyone for joining, encouraging viewers to subscribe on YouTube to stay up to date with Gen AI University.