DSL for prompt

[swerner]

I know its easy to get started with just a string prompt, but I've been thinking an easier to use DSL for prompting might make it possible to do more complicated techniques without writing a long paragraph.

This could also make it easier to do things like RAG, attaching files, audio, video, etc, run different models, insert randomness (Oblique Strategies) into the prompt, or enable essentially a "Prompt Currying" workflow. It also allows us to bake in techniques/deprecate outdated techniques, target techniques to certain models, etc.

In my mind I'm thinking something very similar to what activerecord has with being able to chain methods together that all return the self, but only generate the final result once its needed.

Rough sketch, but something like:

def prompt
  role("Expert programmer in #{@technologies.join(", ")}").
  instructions("write code using #{@technologies.join(", ")}").
  task(@task_description).
  examples(retrieve_task_examples(@task_description)).
  validate_response_with(@task_validator).
  compare_responses_between(:gpt-4o, :claude-3.5, :gemini-1.5-flash).
  use_chain_of_thought
end

[mdeering]

I'm looking at liquid and its Context for this.
https://www.rubydoc.info/gems/liquid/Liquid/Context

Powers Shopify's theme system for theirs store fronts.

[swerner]

Oh interesting! I'm not super familiar with Liquid, I'll look into this, thanks!