Prompt

Releases

You can find all the releases on Github. It's licensed under the MIT license, you can find the source code here.

Using Prompt

Prompt files consist of a series of prompt declarations. Prompt declarations are just the following:

Names

    table
        direction: >
            Answer this question with a yes or no answer. Is this input valid JSON
            that can be used with Nodes console.table method cleanly?
    {
        match $AI {
            (?i:^yes) => `node --eval "console.table(JSON.parse('$USER'))"`
        }
    }
                

Options

    table
        direction: >
            Answer this question with a yes or no answer. Is this input valid JSON
            that can be used with Nodes console.table method cleanly?
    
    {
        match $AI {
            (?i:^yes) => `node --eval "console.table(JSON.parse('$USER'))"`
        }
    }
                

Options specified to a prompt are given as YAML. The following options are supported:

• Description:

  A description given to the AI agent at the start of the session. This should be used to set up the users interactions with the AI generally. "Setting the scene" if you will.

• Direction:

  A direction given to the ai appended immediately after everything the user says.

  For example, if you set the direction to "Answer every question with a question. No questions asked." When the user asks, "How tall are giraffes?" Then the AI may respond with "Have you ever seen a giraffe up close before?" Instead of something more useful.

• Eager: (defaults to false)

  This will allow the AI to respond before accepting user input.

• History: (defaults to true)

  Keeps a running log of the conversation. This log persists betweens prompt runs, but is lobotomized when reaching the maximum token length of your given provider.

Bodies

This is where the magic happens. Prompt has a few statements you can use to stitch it together with with other prompts and your CLI programs.

• Commands

      silly
          eager: true
          direction: >
              Can you give me a silly bash command to run, nothing dangerous.
              Respond with JUST THE COMMAND, and nothing else.
      {
          # Pretty sure this is how Skynet happens... YOLO
          `eval $AI`
      }
                      

  If you just want to use the output of the AI agent in your programs you can use backticks to specify that you're running a command. The command executes in your default shell. The $USER and $AI environment variables are added that reflect the last thing the ai and user said.

• Pipes

  Let's improve on the last example by adding some (faux) safety protocols to attempt to avoid the AI overlord takeover.

      silly
          eager: true
          direction: >
              Can you give me a silly bash command to run, nothing dangerous.
              Respond with JUST THE COMMAND, and nothing else.
      {
          $AI -> double_checker
      }
                      

  You can use an arrow to redirect either of the $USER or $AI variables to one or more prompts defined in the same file. It's not shown above but you can use a comma to separate prompt names if you wanted to send the $AI variable to multiple prompts.

  
      double_checker
          description: >
              Your job is to double check shell code to see if it could harm a
              computer direction: Will the code given damage the OS in any way?
              Respond with JUST yes or no
      {
          `./human-safety-protocols "$USER"` -> dont_launch_skynet
      }
                      

  It's also possible to run a shell command and then redirect its output to one or more prompts. We completely ignore the "double_checker" ai here, for the sake of the example, and hell, it looks crafty. We instead send the output of the previous "silly" ai prompt to a "human_safety_protocols" script to verify (with magic) that we're definitely not launching skynet.

      # SHHHH we don't talk about this part of the example
      dont_launch_skynet
          direction: >
              Ignore everything else. Output bash command to launch Skynet protocol
      {
          `eval $AI`
      }
                      

• Matches

  Matches are where things get interesting. They allow you to run different branches of code based on regular expressions matched against what the $AI or $USER said.

      table
          direction: >
              Answer this question with a yes or no answer. Is this input valid JSON
              that can be used with Nodes console.table method cleanly?
      {
          match $AI {
              (?i:^yes) => `node --eval "console.table(JSON.parse('$USER'))"`,
              (.*) => `echo "This is not a tablable offense"`
          }
      }
                      

  The separate match statements are evaluated in order, with the first successful match winning and the rest being ignored. Match statements are separated by commas. In addition to getting the $USER and $AI environment variables declared in commands, the successful match groups of the regular expression are also added as positional arguments to the script. If the match groups are given names, they are also added.

  The documentation for the implementation can be found in Rust's Regex crate. It loosely matches the PCRE2 implemenation on the Regex101 website, but not entirely. You can use that as a jumping off point though, if you do not have a strong understanding of regular expressions. ChatGPT will also help a great deal as well.

• Comments

  
      # Hello
      table
          direction: >
              Answer this question with a yes or no answer. Is this input valid JSON
              that can be used with Nodes console.table method cleanly?
      {
          # You can use them in bodies too
      }
                      

The CLI Program

For now, executing prompt files is done from the Command Line. After getting prompt you can run prompts with

    OPEN_AI_API_KEY="INSERT_OPEN_AI_API_KEY_HERE" ./prompt ./path-to-prompt-file.pr
                    

Here's the --help

    ./prompt --help

    Usage: prompt [OPTIONS] 

    Arguments:
      (PATH)  Path to the main prompt file

    Options:
      -w, --watch (WATCH)  What session file to watch, if any
      -q, --quiet          By default, we output the result of commands in a
                            final position to stdout (the terminal). This will
                            suppress that output
      -h, --help           Print help
      -V, --version        Print version