I thought I'd write a short post for future reference.

This is how I use LLMs daily for different tasks, such as:

  • Writing a message to a specific audience while taking into account a message history or context (I use Claude Projects for that)
  • Generate ideas to solve a problem—although I try to do that less often to limit my resistance to thinking, which is alarmingly increasing
  • Analyze data (PDFs, spreadsheets)
  • Help me identify reasoning or logical fallacies in my writing
  • Learn about new concepts
  • Investigate trends
  • Answer cooking and household questions
  • Analyze personal health data (privacy hazard, of course)
  • Provide high-level legal advice
  • Perform comprehensive search for a given subject, etc. 

I mostly use o3 and Claude 3.7 Sonnet (Extended Thinking). 

Some heuristics:

  • Keep conversations below 200k tokens (about 50k words). I don't believe in this 1M token thing (I may be irrational)
  • Edit an already sent message instead of trying to tell the AI to correct itself if its response was bad. So let’s say I ask it a question about solar power engineering… it answers in an academic way but I wanted a business approach… instead of sending a new message saying “i want this to be about business”, I edit the first message I sent to steer the AI (less tokens burnt per day, and answers better, as any noise in the context window will prevent optimal behavior)
  • Improve prompts when asking about mission critical stuff (assumption: humans don’t know how to prompt). I use the Anthropic Console for that or I use speech-to-text. I noticed that longer prompts are better and it's easier to speak than type
  • Be diligent about Project context (as little noise as possible, delete old files, clean up regularly)