Day 16 - Prompt Improvement

Here are the basic principles for today:

• LLMs are good at transforming written language.

• We use written language to tell LLMs what to do.

• Therefore, LLMs might be good at transforming simple prompts into better prompts.

The homework for today: start with one, or a few prompts, and ask your LLM to improve them.

The takeaway: you can come up with a few simple prompts, ask the LLM to make them better, then customize the improved prompts and use those as prompts. For this homework, you don’t have to take that last step, but hopefully you’ll see that you could.

Here are three examples, or create your own. Note that these examples each have three prompts embedded within them; for each example, copy the whole thing, with the improvement instruction and the three prompts, to give to your LLM.

Here are three LLM prompts. Make them better:

• Tell me a recipe for bundt cake.

• Give some variations to try.

• What are some other recipes that would go together well?

Here are three LLM prompts. Make them better:

• Write lyrics to a song.

• Critique the lyrics.

• Write an improved version of the song.

Here are three LLM prompts. Make them better:

• Teach me how to write better prompts.

• Help me practice.

• Give me a simple prompt, and I'll try to make it better.

If you like the your lyrics from the second example, you can try using AI to turn them into actual songs with Suno (my fave) or Udio. They both have free tiers.

The third example (teach me how to write better prompts) is a doozy, it can turn into multiple days of homework right there. If you have the time, dive into learning how to write better prompts. Or, put that aside and come back to it bit by bit over the coming weeks.

I tried this “three simple prompts, make them better” technique with ChatGPT 4o, Claude 3.7 Sonnet, and Gemini 2.5 Pro (experimental). They all did a good job, but a little differently, and here is my ranking:

1. Claude - generally did a great job of expansion.

2. Gemini - usually overthought the problem, for example, generating three options for each expansion, and peppering those with commentary and examples. This wasn’t bad, but it meant I had to parse more information to get what I wanted.

3. ChatGPT - did a decent job, but the expansions were simple, less comprehensive than Claude’s.

Try a couple of LLMs, your mileage may vary!

~ Pete