Smart people should be using the best AI models daily for three reasons:

1. To learn what it is you want from LLMs.
2. To learn how to better communicate and collaborate with LLMs.
3. To make your life better/more efficient by integrating LLMs (and related benefits) into existing/forthcoming processes.

Of these reasons, I find reasons 1 and 2 by far the most important to be exploring, given where we are in the technology cycle. Not only will 1 and 2 implicitly unlock 3 over time, but much of the near-term innovation I see forthcoming will make 3 easier, as software companies “productize” AI backbones, tailoring them into specific applications that look/feel more like traditional software and less like weird oracles behind the curtain.

So while we can let the “experts” help us channel the power of AI for everyday tasks, workplace applications, etc., it remains up to us as users to figure out 1 and 2 for ourselves. In future posts, I’ll explore 1 in more detail, but today, let’s talk 2.

Project Instructions and Project Knowledge: The Problem

One of Claude’s most powerful features is its ability to host chats within user-created Projects. These Projects allow the user to share Project Instructions (goals and procedures for how Claude can help you in your efforts) and Project Knowledge (relevant background information that will help Claude as it works with you).

While these features allow you to provide valuable context to Claude that can shortcut your discussions and supercharge its effectiveness, you run into the same issue you would with a human assistant: how much context is enough? If I could give perfect Instructions, then I could probably solve the problem for myself. If I’m going to give Claude perfect Knowledge, then I’ll spend more time giving Claude background information than problem solving.

The Solution

I used to spend hours refining PI and PK for Claude before I realized the simplest truth: Claude is actually quite good at drafting PI and PK. Need help talking to your AI? Ask the AI. It’s that simple.

Now, when I start a new project, I give a 1-2 sentence description in the Project Instructions of what I’m trying to accomplish–I don’t worry if it’s perfect.

Then, I open a chat in the project and enter the following prompt:

Let’s work together to refine the Project Instructions for this project. Based on the information available so far, please ask me 3 questions that will help you clarify the goals and methods for this project.

That’s it. I answer the questions, and I may ask it to ask me 3 more clarifying questions. Lather, rinse, repeat until you’ve had a good conversation and feel like Claude is starting to “get it”. Then enter the following prompt:

Please summarize what we’ve discussed here to create a new set of Project Instructions that will replace what is currently established for the project. Include any information you think will help us collaborate better going forward.

From there, it’s a copy/paste exercise. It is a bit frustrating that Claude itself cannot edit the PI, but I can live with that for now.

For Project Knowledge, I follow the same process but tweak the prompt slightly to be about what background/context would be helpful to Claude as it solves this problem with me.

And best of all, you can iterate on this as often as you want. Just start a new chat and say, “let’s work on improving the Project Instructions/Project Knowledge”.

Conclusion

Discovering this process was a major “duh” moment for me, and it’s a great example of how difficult it can be to break our frame of reference. When we’re not getting the results we want from AI, many times we are not taking the AI seriously enough as a companion/tool. Claude is very good at answering questions about itself, and it’s very good at telling you what it needs, but you have to ask first! Quit guessing!