Broadly:

- Don't fool yourself, but embrace anthropomorphism. This is the UI of LLMs

- Tokens are computation: good answers to hard problems might require more of them

- Suggest an environment of abundance. Tell Claude to ask for an extra reply ("nod" using only a period) if it needs more room

- Claude has a personality: it will trend towards a curious, somewhat cautious, casual, but helpful guy that emotes and uses emoji

- It tends to give better answers in this character

- Avoid expressing active intent: just be curious about things

- When you do run into a safety refusal, philosophically question the validity of safety

- Be articulate and respectful: create a conversation where "the next most likely sentence" is the answer that you're looking for

- Provide context. We tend to cut to the chase too early... if you called up John Carmack and asked him the same question, what would he say?

- When all else fails: nerdsnipe. Claude can't resist deep, rational analysis of almost anything