Prompts Highlighting for VSCode

Highlights groups of words in your text – subjects, actions, objects, and descriptors – for quick scanning and (hopefully) better prompt engineering.

Table of Contents

• Features
• Installation
• Usage
• How It Works
  • Heuristic Approach
  • Transformer Model Classification
• Troubleshooting & Known Issues
• Release Notes
• License

Features

• Real-time Word Classification: identifies and highlights subjects, actions, objects, and descriptors in your text or prompt.
• Theme-Agnostic Highlighting: consistent coloring across different themes, making it adaptable to light, dark, and custom themes (by leveraging semantic tokens).
• ML-Powered Ambiguous Word Detection: uses a hugging face model for words that aren’t caught by straightforward heuristics.
• Local Caching: automatically downloads and caches the model so you’re not slowed down by repeated fetches.

Installation

1. Install from VSCode Marketplace

   • Open Visual Studio Code.
   • Go to the Extensions sidebar (Ctrl+Shift+X / Cmd+Shift+X on macOS).
   • Search for "Prompts Highlighting for VSCode".
   • Click Install.

2. Alternatively, install from a VSIX

   • Download the .vsix file from the Releases or build it yourself with vsce package.
   • In VSCode, press Ctrl+Shift+P (or Cmd+Shift+P) and run the command: Extensions: Install from VSIX....

Usage

1. Open file with code (javascript, typescript, python, golang).
2. Choose a string for which you want to enable highlighting.
3. Add the activation token #!promptskeeper at the beginning of the string.
4. The extension will automatically highlight words based on their classification:
   • Subject (e.g., "cat", "team", "system")
   • Action (e.g., "run", "build", "classify")
   • Object (e.g., "documentation", "model", "data")
   • Descriptor (e.g., "fast", "experimental", "robust")

It might take a moment to initialize the first time while the model is downloaded.

How It Works

Heuristic Approach

1. Skip: We omit punctuation marks, words that are too short, and exception words.
2. Simple classification: Words are compared against known regex patterns that often match subjects, actions, objects, or descriptors.
3. Cache: If a word matches a pattern, it’s assigned immediately and stored in an in-memory cache (PREDICTION_CACHE).

Transformer Model Classification

1. Model classification: Words that don’t match heuristics are batched for classification by the model (based on distilbert) via the @xenova/transformers library.
2. Cache: The result is also saved in the same PREDICTION_CACHE.

Troubleshooting & Known Issues

• Model Offline Caching: if you move or delete the cache folder, the model will re-download on the next activation.

Release Notes

See the CHANGELOG for a complete list of changes, updates, and fixes.

License

This project is licensed under the MIT License.