3 sessions inside Boa
Jeremy Cabral
Co-founder, Finder
Jeremy is a co-founder of Finder, the Australian financial comparison site used by over 35 million people across Australia, the US, UK and Canada. He was Chief Operating Officer, running product, design, tech, growth and international expansion into more than 20 countries, and helped bootstrap the business past $100 million in revenue before it raised $30 million at the end of December 2021. He has since stepped out of his operational role, remaining a special advisor to the board, and now advises businesses on AI growth strategy and operations.
The sessions
What Jeremy teaches
Stop single-shot prompting. Ground the model with project files.
Cabral says a single prompt gives generic output that everyone can spot. Upload files into ChatGPT (or Claude) and use the projects feature so the same reference material is pulled into every response going forward. Add custom instructions for tone, target audience and Australian English spelling, because American Zs in your copy are an obvious tell that it came from ChatGPT.
From Jeremy Cabral: AI Agents, SEO and Growth Engines →Ask the tool to write your prompt template first
Instead of guessing at structure, tell the model your goal: you want a project file that informs all future responses, so give me a prompt template to fill out. Fill it in, load it back into the project, and the LLM has structured knowledge it can reference easily. Unstructured files take more effort for the tool to use.
From Jeremy Cabral: AI Agents, SEO and Growth Engines →Shop between models for each task
Cabral used Gemini to write the prompt for Leonardo AI (the image tool Canva acquired), then ran that prompt through ChatGPT because the result was better. Gemini produced the best prompt structure and the most repeatable image outputs. Models change constantly, so test the dropdown rather than defaulting to one tool.
From Jeremy Cabral: AI Agents, SEO and Growth Engines →Use reasoning models for multi-step SEO work
For website tasks, Cabral feeds O3 three inputs: a crawl of your site from a tool like Screaming Frog, your target keywords, and competitor URLs. Add a clear structured goal and it will execute the task. Expect 10 to 15 minutes or longer per run, not a seconds-long answer.
From Jeremy Cabral: AI Agents, SEO and Growth Engines →Build the four foundations before you touch an AI tool
Jeremy asked the room to commit to four things: one central spreadsheet holding every project (a solopreneur can start there), all documents in one editable place rather than email as the source of truth, transcribing every meeting after asking permission, and hiring a virtual assistant. The VA matters even in a tiny business because it forces you to break work into repeatable, delegable processes, which is exactly what an AI agent needs later.
From Building AI systems that actually scale your business →Ask ChatGPT to write the prompt you then paste into a fresh chat
Most people write a one-line zero-shot prompt and get a generic answer customers can spot immediately. Instead, ask ChatGPT to act as a prompt engineer for your task, take the prompt it produces and paste it into a new chat. Jeremy called this the single tip to take away if you take nothing else.
From Building AI systems that actually scale your business →Load a project file with your pricing, products and FAQs
Anything you repeatedly refer to when answering clients or drafting documents should live in a ChatGPT project file, whether or not it is public on your website. That context stops the model producing generic output. For his marathon Jeremy loaded roughly ten documents into a project and prompted ChatGPT 3,049 times across preparation and the race itself, with the quality of answers coming from the stored context.
From Building AI systems that actually scale your business →Switch off Auto and run Thinking as your default model
ChatGPT defaults to automatic model selection, and Jeremy's read is that OpenAI's cash burn means you are not always getting the best model for the question. Set it to Thinking for most work, use Pro for specific tasks if you have the account, and turn Thinking off mid-response when you just need something fast. Test the same prompt across models in both ChatGPT and Claude and compare.
From Building AI systems that actually scale your business →Treat any ChatGPT draft as a 1 out of 10 and build from there
Jeremy's point is that AI output is probably better than most of us could write, which is exactly the problem. If it takes one prompt, everyone has it. Ask what would make the same page a 10 out of 10, then do that work. Never publish straight from ChatGPT to the internet.
From Finder's Jeremy Cabral on SEO, AI search and Reddit →Print the page and read it aloud before you publish
A trick Finder used for years. Print out the content and read it out loud. If the words are embarrassing to say, the content is crappy. On screen you skim and skip sentences, so you miss the fluff. Then hand it to someone and ask directly whether it is useful.
From Finder's Jeremy Cabral on SEO, AI search and Reddit →Build 6 to 8 guides as a cluster before moving to the next topic
Do not write whatever interests you. Pick one topic, map the 6 to 8 guides your customer actually needs, link them to each other and keep overlap between them low. Each guide answers a distinct intent. The cluster has to be logical for your customers and make sense to Google.
From Finder's Jeremy Cabral on SEO, AI search and Reddit →Grow three separate kinds of authority, not just one
Topical authority is how deep and consistently you cover a subject. Domain authority comes from links in, then distributing those links across the site so pages get crawled. Entity authority is Google's knowledge graph understanding of your company and your authors, so publish qualifications and get your people quoted and referenced elsewhere on the internet.
From Finder's Jeremy Cabral on SEO, AI search and Reddit →
Watch the full sessioninside Boa.
Session notes are free. The recordings, the live advisories and the founders in the room are for members.