I've been trying to figure out a framework that I could repeatedly use in order to research deeply on any topic of interest. This is an actively updated piece with practices that I've found helpful over time. It's very rough, so expect it to change over time. What works most is that you should not let LLMs come up with their own narratives. You should think of them as highly capable data mutators. I spend most of my time curating the data that I will be using for my research. > 'Research' here is broad. It can mean anything from looking at China's energy usage to AI compute bottlenecks. Work really hard to find good sources. LLMs are still bad at this. I rely on my X feed, substack, YouTube, and websites like [Our World In Data](https://ourworldindata.org/) Once you have a solid, reliable corpus dive deep into it with coding agents. Use Claude Code or Codex to shape the data in a way that is most intuitive to you. Ask agents to create graphs, charts, tables grounded in the corpus you define. And then your job collapses to that of asking good questions. Ask good questions and find answers to them. Do not ask the agents what they think. This usually does not work well. I hope to publish some of my workflows over time as I find out what works.