Writing Prompts
In a lot of my work with AI tools and agents, I’ve spent considerable time writing prompts and iterating over them. A decent prompt can get you 80% of the way there, but then it takes a long time to cover the last 20%. And while in smaller tasks the last 20% can be ok, for larger tasks the last 20% is still something the industry is struggling with.
So in my journey in getting better with prompts, I have found the following few resources to be helpful.
Fabric
I have followed Daniel Miessler for a little bit, and I found that he has a repository of prompts called Fabric. I have found it to be a great resource for finding prompts for a lot of different tasks.
When I start thinking about writing a task or a system prompt, I look at this for examples to get started with.
Gemini
Gemini has focused on helping organizations get better at using AI with their Vertex AI offering. They have published incredible resources for writing prompts for their models.
Here is the one from them recently which is pretty good.
Claude
Prompt Library from claude is pretty good too. It has a lot of good examples that are worth going through, if you are starting out.
Final Thoughts
Other LLM providers also have their resources and you can find them on their websites. For me, I have found success in a “prompt writer” that I have set up with Gemini. I found that as humans it’s harder for us to be very verbose in writing prompts that work well with LLMs. So I ended up writing a prompt in Gemini whose task is to write prompts for other agents. Then the writer prompt spits out, in steps and a verbose manner, what I am looking to achieve from agents. It mostly ends up being a good starting point. Then I feed the prompt writer the changes I want, based on agent’s performance and we iterate.
This video below also has some really good ideas on how to write prompts and in general how to work with AI agents.
