"AI: High Effort/Low Effort Use" Topic

I've been enjoying a lot of success with AI tools recently, both writing apps and analyzing game situations, as well as doing other kinds of research. It's been amazing how quickly I can get good results.

Meanwhile, here and on boardgamegeek.com and other places I see people posting the results of their AI experiments and they're pretty weak tea. Mistakes or hallucinations, shallow understanding of the questions, etc.

I'm also reading a novel written with the help of AI. Never mind why I'm making myself read it, but I am. It's what you would expect.

I have a guess at what's causing this disparity, why there is slop and why I am getting much better results personally, and I wanted to share it in case it helps. When I prompt Claude (or ChatGPT) I write LONG prompts. I take 20 minutes to write five hundred words, even up to an hour to write a thousand and one words. I include pictures. I specify where I want to go, how I'm going to be looking at or using the answer. I give example output if possible. I don't think these little essays need to be perfectly organized, and it's okay if they are a little repetitive – those would be bad if you were writing for a human, but they LLM doesn't care. But there is a lot of specific info in my prompts about what I want.

I also iterate. If the app doesn't look like I want it to, if the answer to a question is unsatisfactory in any way, I describe the problem and have it take another pass.

I suspect the other folks who have gotten unimpressive output have written very short prompts. They don't talk a lot about their process, but what they do say is consistent with a lower effort prompt and no one mentions iteration.

So if you're going to use AI for something – and I would encourage it, it's a powerful tool! – spend some time thinking about what you really want, all the elements you want taken into consideration, what the output should include, how you're going to use it, and even the things you want left out.

Garbage In. Garbage Out. As ever.

Good luck and have fun!

Yes, that is exactly my experience too I use ChatGPT and Grok pretty interchangeably and often use the same long and carefully written prompt with each. Results will be 90% similar and the differences are often interesting

I would add one thing to your excellent advice. If you ask the AI to tell you what it needs and how best to phrase it, the robot will do that. I posted here but I don't think many noticed, but when I was trying to fit my dead guys inside a 2" base, Grok told me that its image maker will not recognize 28mm as a measurement. If the disk is 2 inches and I tell it to make the image head to toe one inch, it can do that So part of doing prompts is prompting the AI to suggest how to write prompts.

Grok told me that the default is either to make the image fill the space from rim to rim, or else to make it small in the center. It gave me explicit language to include to stop it from doing that.

The AI knows itself it will reveal "secrets" that are just assumptions or SOPs if you describe the problem you are having with its answers and ask for help. A year ago I could not get Chat to do a dead Indian though it had no problem with a dead ranger. I pointed out this inconsistency and it admitted it and said my project should be an exception but could not override the instruction. But if could do a SLEEPING Indian if I asked for one which we did. Happily that woke idiocy no longer holds true