I’m using Chats GPT to build a swim training session. It does some great sessions which is fine. But no matter how I tell it to match the total distance I input, it never comes out right. So if I say to write a 2000m set, it will make it 2300m or something along those lines. Here’s is my chatbot prompt with a sample input. There’s other issues with it too but I want to get the total session distance right first. I’m using the ChatGPT 3.5 Turbo with a temp of .5

I will send this info: : Total Distance of session in meters, Length of pool in meters, pace per 100m freestyle, and focus of the session… Please create a swimming session with warmup, kick, pull, drills, a mainset, and a cooldown. Put it in table format but do not number the sets. Column A, Description of the set, B, reps and distance per each like (4 x 25m) and and interval, C Total time of set, and D, total distance of set. (If the row is just a description don’t put a total distance for D in that row). At the bottom of C add up the total time. And at the bottom of D add up the total distance of the sets. The sets must be created so the total of all of them is exactly the same as the request total distance in meters.

You could try a multi-step approach and ask ChatGPT to reflect on its plan. Try assembling a prompt with the following structure:

draft a swim training session in <draft> tags meeting conditions X,Y,Z

reflect on the draft you assembled. Confirm that it does, indeed, meet these conditions.

return a finalized plan in <plan> tags, making any amendments to your <draft> to make sure it meets the conditions.

I can’t say for sure that it will work, but this kind of prompting does improve performance on problems requiring multi-step reasoning quite a lot in my experience.

I gave it a couple examples but it seems to have come out worse!

So if I start but giving you 3000m in my initial total session distance entry, all the sets should add up in a total session distance to exactly 3000m. And if I give 3500m in my iniital total distance entry, all the sets should add up to a total session distance of exactly 3500m

Yes, I must say I have used the GPT a lot for so many things and gently refining it as I go seems to work best. But I’m sort of after a one-stop session generator, which I’m not sure is possible. I enter my parameters and want it to spit out a perfect result! I’m hoping to not have to coax it along all the time.

I mean to try putting all of that in one prompt—so it won’t be a series of prompts, refining it at each step; it’s asking ChatGPT to add a reflection/revision step to its (single) response.