Sublime

Here is the single most powerful prompting super-power that I wish the people building AI agents understood: (This is the secret that makes https://t.co/Suo6Df4QyV work SO well and never repeat itself and always be creative, even in GPT-3.5-Turbo mode) The most important thing to remember about GPT/LLMs, in general, is that they are web-page-auto-completers above all, even when they're fine-tuned into assistants, the core of their understanding of the world is built around matching the style & content of their existing prompt as if they were finishing a blog post. So the most effective way to get what you want isn't to ask for it, it's to assemble a blog post about HOW YOU GOT THE RESULT YOU WANTED (Key: past tense) Most factual information on the internet, even the surprising stuff is in past tense. Only fiction and news is present tense, and we all know fiction and news is bullshit. So write your prompts in the past tense. And then there's absolutely NO reason your blog post should end with conversation history. No real blog post about a crazy AI outcome would end with a conversation history. That's why your model loses coherence, makes worse suggestions in longer outputs, and loses track of its core goal and next tasks over time. So write your prompt in a way that *highlights* the insane outcome you're looking for from your agent as if it already happened, and put the long and short-term conversation history somewhere else higher up in the prompt, and at the *very* end of your prompt remind GPT *exactly* what the next output should be and why. For agents, this means your next task and current long term goal should be shown in the prompt *right before* the next output from the LLM, not a history of its recent actions. I'll give you an outline of how we do this and concrete examples in screenshots. For example, here's the breakdown of headings and sections in https://t.co/Suo6Df4QyV... See more