While conducting deep research on a new domain recently, there was one specific friction I encountered frequently: AI appears to do some research, read some websites, and spits out a bunch of statements. Some are true, some are inferred, and some are false. The only way I had to determine he validity of the statements came down to researching each one further, or retesting the points back against the AI, to which I usually get the classic "sorry, my statement was not true".

This happens because (from my observations), LLMs like to take 3 methods to research:

1. Search the internet for sources
2. Refer to it's own training data
3. Form it's own observations based on the context

The critical issue when you need research accuracy is: the LLM rarely indicates which method it derives its statements from. So something that appears as fact could in fact be the LLM's guess.

These days, anyone who boasts about having 100 agents working overnight to build their whatever, for me, just means they've 100x'd their sloppy output.

AI is great at getting you to 80%. As the 80-20 rule goes - 80% is the good enough/gets the job done threshold! But for those of us who aspire to go beyond good enough, myself included - refining something to hit 90-100% of the desired outcome is, in my opinion, just as hard as it was without AI (great to know I still have something to do).

Why is this part hard? Well, getting to 80% is easy because that covers the general building blocks, and AI is good at handling the general stuff. That other 20% though, is your unique vision, and until AI can read minds, we need to direct it to execute our vision - this takes precision, and a lot of back-and-forth.

I worked on a project recently where I had to go very heavy on research, which led to the development of this research prompt. The process became very frustrating as I was closing in on a deadline. Whenever I ask a question expecting a short answer - I get an essay back instead.

AI knows everything and has a mandate to serve us.

After the storm, I reflected on this process, and concluded the above, which now informs how I approach managing an AI, which I'll cover on the next slide.

I bought a 300-page trading textbook, but it makes for a difficult read. What mattered more was being able to reference specific sections when needed, rather than read everything in one go. So I decided to virtualise the textbook for myself - and just ask the author questions directly instead.

The obvious move was to hand Claude the book contents and let it answer. However, the author had a certain way with his words and concept framing, so I wanted to make the knowledge accessible as a conversation - like I'm talking to the author, rather than searching a document.

Most of the work, it turned out, was reshaping the source itself.

For years, Apple notes was where I would log ideas, meeting notes, recipes, todos etc. 1000 notes later, a lot of them have disappeared into the void. For most notes, this is probably fine. But I've always wanted a better system. In the past, I have looked into "second brain" setups, but I lack the discipline to maintain it - it takes a lot of effort. This is where Obsidian MCP + Claude changes the game. I created 4 Claude skills, scheduled to run daily to organise my notes.

If you are not familiar with Obsidian, it's a note-taking tool that has been around for a while. But what makes it great is that all the notes are saved locally as markdown files, it has an amazing library of community built plugins that allows you to customise everything to suit your workflows, and this is also what makes it so great when paired with AI. However, I do think it's better for those who want to tinker with their notes app - so skip if you're not the tinkering type.