PUDL Agent Skills Feedback

[zaneselvans]

• We've created a couple of Agent Skills to help users navigate the hundreds of tables and thousands of columns of energy data that we publish through the PUDL project.
• These skills are agent-agnostic and should work with both proprietary and open weights models in a variety of harnesses. You'll need to refer to the documentation for your particular LLM tooling to see how to install them. (e.g. Claude or OpenCode)
• The skills are primarily focused on working with the PUDL metadata and that's where we've focused our internal testing. So think of them as an adjunct to our metadata search and data preview tool
• ⚠️ Like all AI tooling they can be unreliable at times so double check the answers if you're working on something important!
• Exact behavior and performance will depend on the underlying model and harness you're using, as well as the PUDL metadata and skill.

Share your feedback

• Please use this Discussion to share feedback on the skills!
• Let us know if they go off the rails, or help you do something cool.
• If you have a specific bug to report or feature request, feel free to open an issue

Links

• Catalyst Agent Skills Documentation (definitely a work in progress!).
• The Catalyst Agent Skills repository.
• Background on Agents Skills more generally.
• Tips from Simon Willison on using agents for data analysis

[krivard]

If anyone is looking for prompt ideas, we threw a bunch of them in an evaluation issue: catalyst-cooperative/agent-skills#5