promptgen

CLI for managing and generating Foundation Model prompts

Structure

Prompts can be stored under src\prompts in hierarchical form for the target LLM (e.g. bing, GPT4, Midjourney). Prompts are written in a yaml file which makes it easy to store both the "prompt code" and metadata about that prompt, using the following format.

Prompt Format

The Prompt yaml format contains a few key components:

• name
  • Symbolic name allowing for easy reference from other prompts
• template
  • Template text for the prompt, written in handlebarsjs format
• description
  • Description for the user, explaining why they might want to use this prompt
• use-when
  • Description for the LLM, useful for skills-based prompts (e.g. if GPT4 needs to use Midjourney to do its work)
• inputs
  • Array of expected inputs to the prompt template (e.g. a research prompt might take a topic). Note currently that no type information is included.
• is-partial
  • Boolean for whether this
• references
  • Array of references to other prompt templates (typically partials) that are then loaded and stored as partials so they can be referenced in the handlebars template.

Notes:

Handlebars partial references should use the name of the referred-to template (e.g. bing/accurate has a name of Accurate so is referred to as {{>Accurate}} in the handlebars template).

The separation of references and use of name in this way allows simple swapping of behaviors - you can have multiple experiments with different ways of eliciting "accurate" behavior, and as long as all partials share the same name merely swapping the references will allow you to use the same handlebars template.

The goal is to allow command-line overriding of the references to make that sort of experimentation easier.

Contributing

This project welcomes contributions and suggestions. Although this project is not (yet) under the Microsoft OSS banner, it has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact opencode@microsoft.com with any additional questions or comments.

NOTE on the code - this is my first TypeScript project since before it was called TypeScript, and my first Node.js project in years. I've been using GitHub Copilot X to help me re-learn the language (from my now-native Python), but if you see areas where I should be doing things differently or have corrections for me - I'm quite happy to hear them either in the form of PR or GitHub Issues.

Usage

Usage: promptgen [options] [command]

CLI for managing and generating Foundation Model prompts

Options:
  -V, --version                        output the version number
  -d, --debug                          Enable debug logging
  -h, --help                           display help for command

Commands:
  prompt [options] <string> [data...]  Generate the given prompt
  help [command]                       display help for command


Examples:
  $ promptgen prompt bing/research topic bugs
  $ promptgen prompt gpt/challenge-network -f challenge-personas
  $ promptgen prompt midjourney/gpt-to-mj -f midjourney-examples

Usage: promptgen prompt [options] <string> [data...]

Generate the given prompt

Arguments:
  string                    prompt name
  data                      Key value pairs of data to bind to the template

Options:
  -f, --file <path>         YAML file containing template data bindings
                            (directory assumed to be ./bindings, .yaml
                            extension assumed)
  -r, --prompt-root <path>  Root path for prompts (default: "./src/prompts")
  -h, --help                display help for command