YouTube is full of breathless videos on the new capabilities of Claude, Gemini and ChatGPT every day. There’s nothing wrong with all of them, but ultimately just try it and see what happens. To prevent getting emotionally invested into that solution-in-search-of-a-problem, start with your business problem first.
In my case I wanted to do intensive research on technologies widely used in my industry. In past research projects I’ve used NotebookLM with great success. This time I tried Claude.
I started with a question for Claude and ended with a stunning new workflow for deep research on any topic. It was automatically tailored to my writing style and research priorities. It tests for bias, comprehension, and usefulness in my daily professional life.
Mind = blown. Here’s how it went.
After uploading my context vault – a file containing facts about my background, professional goals, and preferred communication styles – I told Claude I wanted comprehensive research that I can refer back to as a knowledge base and as a tool for uncovering trends and advising clients.
In the casual tone that we all use at the start of a meeting, Claude confirmed the scope of my research and proposed seven subtopics, each one in sequential order and supporting the previous one. Because I requested in-depth responses, I directed Claude to ask me comprehension questions on key areas and follow-up details.
So far we’ve done four of the seven sections. In each one, Claude took less than 30 seconds to develop a richly detailed and well-cited report while I took up to an hour thinking about it and writing a response. It was like the back-and-forth study session in college when you and your study partners realize you’re actually making progress.
Hallucinations Are Real
Before long I saw that while the reports were well-written and thorough, the citations were wrong. Sloppy. One even linked to a professional soccer team on Wikipedia.
As your AI threads get very long, the risk increases that AI will drop previous bits of data, misdirect facts, or sometimes make stuff up. In other words, AI can be surprisingly human, just a lot faster.
Claude confidently stated that a certain concept was relevant to a client project I had in Texas. In fact, that project was in Arizona. When I called this out, Claude sheepishly replied that “I incorrectly attributed the Texas example to your personal experience. That was a fabrication on my part, which is exactly the kind of thing you should push back on.”
Does that mean we forget about AI? No. It does mean that we check AI’s work, follow the citations, ask to restate items, and ask for more specific details. Know your AI’s limits. Know when to start a new thread and how to organize content for clearing up more space.
Despite occasional errors, I have developed a rich knowledge base covering on technology in my industry that, with fact-checking and refinements, I can use in several ways including:
- A personal repository of knowledge that I just keep adding to.
- Background detail for advising clients on industry trends, risks, profitability, and the business problems keeping them awake at night.
- Historical review that I can expand and add to in any direction. I can test it for bias and ask it to revisit the same information from a different perspective.
Don’t forget the big caveat: No one will care that it was your AI’s errors, and not you, that got someone in trouble, sued, fired, or worse. Your name and your reputation are on the line when you’re advancing AI-driven material.
Automate your workflows, but figure out what they are first
Three more topics remain for my research. I won’t go further in that first thread because I want fresh new memory. But first, I turned the entire thread into a Claude Skill.
Skills in Claude can be thought of as workflows, recipes, or a set of sequential tasks. Each task is its own prompt in plain language. Together they are highly sophisticated and forehead-slappingly brilliant.
At the end of my thread, I dropped in this prompt:
Make a skill called “deep-topic-research” which creates a formatted research structure and then conducts web-based research and leads me in dialogue and comprehension review based on the methodology that was created and used in this thread. Before beginning, we identify what the successful result looks like – the completed research – based on each of our mutual suggestions. The emphasis of this skill is on developing a reliable, readable, and documented body of knowledge. All sources should include citations.
The resulting three-page bulleted outline of sophisticated research tasks was stunning. AI had taught itself to make a complex new workflow just from chatting with me. I can revisit it and make improvements. Claude’s parting words after completing this skill:
“One suggestion for a future iteration: once you have used the skill on a second topic outside of this one, we can refine it based on what transfers well and what needs adjustment. Skills improve with real use.”
In the end, AI is best for helping you put your thoughts together. What you do with them next is up to you.
