I love the framework, but that's a lot to follow and remember for people who engage constantly. With those principles and steps in mind, a shortcut is to think of yourself as having a 2 way conversation with a subordinate. or colleague....lots of guidance, back and forth, iterative conversation.
Rules with the reason is the tip that carries straight into research prompts. 'Flag any quarter where operating cash flow trails net income, because that gap is what I underwrite' catches cases a bare 'check cash flow' misses.
Rules with the reason attached is the tip I'd put first. "Keep it under 150 words, I'll read it on my phone" gets better results than any amount of capital letters. Same for image briefs: say why the background is white, not only that it is.
This mirrors what's happening with agent frameworks generally, not just Claude prompts. As models hold context and run autonomously for hours, the bottleneck moves from 'how do I phrase this instruction' to 'how do I define done.' Curious whether you've noticed Claude handling ambiguous success criteria differently than it did a year ago, or if it still needs that spelled out explicitly.
Both habits came out of papers. 'Let's think step by step' took InstructGPT from 17.7% to 78.7% on MultiArith in 2022, and 'Take a deep breath' was the best prompt DeepMind's OPRO found for PaLM 2 a year later.
I love the framework, but that's a lot to follow and remember for people who engage constantly. With those principles and steps in mind, a shortcut is to think of yourself as having a 2 way conversation with a subordinate. or colleague....lots of guidance, back and forth, iterative conversation.
Rules with the reason is the tip that carries straight into research prompts. 'Flag any quarter where operating cash flow trails net income, because that gap is what I underwrite' catches cases a bare 'check cash flow' misses.
Rules with the reason attached is the tip I'd put first. "Keep it under 150 words, I'll read it on my phone" gets better results than any amount of capital letters. Same for image briefs: say why the background is white, not only that it is.
This mirrors what's happening with agent frameworks generally, not just Claude prompts. As models hold context and run autonomously for hours, the bottleneck moves from 'how do I phrase this instruction' to 'how do I define done.' Curious whether you've noticed Claude handling ambiguous success criteria differently than it did a year ago, or if it still needs that spelled out explicitly.
Both habits came out of papers. 'Let's think step by step' took InstructGPT from 17.7% to 78.7% on MultiArith in 2022, and 'Take a deep breath' was the best prompt DeepMind's OPRO found for PaLM 2 a year later.
There's a book on Amazon that goes in depth on this: "The Prompt Master’s Guide to the Galaxy"