Put your Claude Project on steroids: A small tip as an addition to Brandon's 'New Claude 3.5 Crash Course' video

Aug '24 • Full Stack Development with AI

I've been working on a complex project with a GraphDB ontology and found a helpful approach. Setting a system prompt for Claude's Project has significantly improved my workflow. I use Claude's console to generate this prompt, which I then add to the project.

Here's a link to check out the two templates: one is the SYSTEM_PROMPT for Claude's Project, the other is a template for a PROJECT_BLUEPRINT.

https://gist.github.com/bartolli/5291b7dd4940b04903cae6141303b50d

Here's the workflow I use to generate these:

1. I import the project requirements or brief into Claude and ask to refine and adjust it to my liking.

2. Once I am happy with the brief, I ask Claude to generate a system prompt.

3. I copy the result when I am satisfied, then I move to Claude Console and go to Generate Prompt, paste this and get the response back.

4. The output is my system prompt file. Run this prompt in the console for a workbench evaluation and fine-tuning, adjust to your liking. Copy the prompt.

5. I go back to my Claude Project and click Edit on the right-hand panel, right below the progress bar of the files storage where it says "Custom Instructions will change how Claude behaves and ..." - I paste my system prompt here.

6. I upload the PROJECT_BLUEPRINT.

It's important to mention that using the XML tags and curly brackets somehow made Claude respond faster. I assume that those tags give Claude direct access to the uploaded artifacts without having to perform any RAG operations, which is also token-friendly.

A key feature of this prompt is how it guides Claude in choosing appropriate artifact types. It instructs Claude to use Markdown tables for simple comparisons, LaTeX for complex tabular data, Mermaid diagrams for processes and relationships, code blocks for executable code, and SVG for custom graphics. This ensures that information is presented in the most effective and clear manner.

Now you're ready to start querying Claude.

I made a few experiments today trying to ask Claude to help me build a car. I was rolling on the floor laughing. You should try this out and post the response here! :))

This method helps me avoid token timeouts and saves tokens by instructing Claude to return only relevant code blocks instead of entire files. It's improved the quality and focus of responses while streamlining the process.

Brandon's full crash course video here:

https://www.skool.com/ai-developer-accelerator/new-claude-35-crash-course

Best,

Angel

12 comments