Command line tool for creating LLM prompts from your code using Handlebars templates.
This was a project to brush up on Rust and is based on code2prompt with some additional functionality that I found useful.
Manually copy and pasting code and code snippets to LLMs has several issues:
- You lose formatting and structure. For a language like Python that uses whitespace, the formatting and structure of the code is important for the LLM to accurately and comprehensively understand the code for your request.
- You lose context. In order for the LLM to provide the best responses, it requires the context from all of your code. This includes your code snippet's dependencies, other user defined modules/objects, and other scope variables and concepts.
- You lose prompt standardization: If you are using LLMs to perform repeated tasks like documenting code, you have to make sure that your documentation styling guidelines are consistent across every request.
To get the most out of LLMs, prompting has to be clear, comprehensive, and consistent. How code prompts fixes these issues:
- Code prompts will ensure the inherent structure of your code is retained to limit any possible misinterpretation or inferrence required from the LLM.
- The complete context required for understanding the code is enclosed in a standardized format. This includes the project directory tree and file pathing.
- By using Handlebar templates, the prompts will be comprehensive and consistent. If your preferred style guide and/or coding practices are updated in the future, there is no need to remember to update all future manually typed prompts. Instead, just update the target Handlebars template once.
To download and use the code prompts command-line tool, you have two options: you can download the release binary or compile from source. For more detailed steps, including setting up tab completions, see the setup guide.
To download a release binary, go to the releases and download the binary for your OS.
To build from source you will need to have git, Rust, and Cargo (will be installed with Rust) installed.
First clone the repository:
git clone [email protected]:seankim658/codeprompts.git
And then compile a release binary:
cd codeprompts/
cargo build --release
More detailed usage guides can be found here.
More extensive documentation on the options can be found here. The code prompts command line tool has the following arguments:
Create standardized LLM prompts from your code
Usage: codeprompt [OPTIONS] [PATH] [COMMAND]
Commands:
completion Generate shell completion scripts.
help Print this message or the help of the given subcommand(s)
Arguments:
[PATH] Path to project directory
Options:
--include <INCLUDE> Glob patterns to include
--exclude <EXCLUDE> Glob patterns to exclude
--exclude-priority Change pattern priority in case of conflict to prioritize the exclusion pattern
--exclude-from-tree Eclude files/folders from the source tree based on exclude patterns
--gitignore Don't respect .gitignore file
-d, --diff-staged Capture the git diff for staged changes only (equivalent to running `git diff --cached` or `git diff --staged`
-u, --diff-unstaged Capture the git diff for unstaged changes only (equivalent to running `git diff`)
--no-tokens Don't display approximate token count of the genrated prompt
-c, --encoding <ENCODING> Tokenizer to use for token count [default: cl100k]
-o, --output <OUTPUT> Redirect output to file
-l, --no-line-numbers Turn off line numbers in source code blocks
--no-codeblock Disable wrapping code inside markdown code blocks
--relative-paths Use relative paths instead of absolute paths, including parent directory
--no-clipboard Disable copying to clipboard
-t, --template <TEMPLATE> Optional path to Handlebars template
--no-spinner Whether to render the spinner
--json Whether to print the output as JSON. Defaults to False
--issue <ISSUE> Fetch a specific Github issue for the repository
--verbose Run in verbose mode to investigate glob pattern matching
-h, --help Print help (see more with '--help')
-V, --version Print version
The templates use a simple templating language called Handlebars.
The pre-defined templates can be downloaded from the project releases. Download the templates.zip
.
Currently, the included pre-defined templates are:
Template Name | Description |
---|---|
default_template.hbs |
This is a simple default template that will structure your project path, source tree, and code blocks. |
documentation_template.hbs |
The documentation template creates a prompt for documenting code. The documentation guidelines are consistent with the HIVE lab guidelines and documentation requirements. |
git_commit.hbs |
Template for creating a concise and accurate git commit message. Can be used with both the diff-staged and diff-unstaged options. |
git_issues.hbs |
Template for implementing changes based on a Github issue. |
code_optimization.hbs |
Template for optimizing code in time and space complexity. |