Get the weekly radar
One short email with the freshest products, every week.
Vibe coding tool with prompts in the repo
Grounded in available product and source data
A self-compiling tool that transforms prompt stacks into code and tests using LLM generation.
vibec is a unique compiler that processes markdown-based prompt stacks to generate code, tests, and documentation. It can compile itself through a bootstrap process, evolving its own implementation (bin/vibec.js) across numbered stages. The tool supports both static (.md) and dynamic (.js) plugins, maintains staged outputs in output/stacks/ for Git history, and aggregates the latest runtime version in output/current/ using a "Last-Wins" merge strategy.
vibec/
├── bin/ # Initial implementation
│ ├── vibec.js # Core compiler script
│ └── test.sh # Test runner
├── bootstrap/ # Bootstrap documentation
├── stacks/ # Prompt stacks
│ ├── core/ # Core functionality
│ │ ├── 001_add_logging.md
│ │ ├── 002_add_plugins.md
│ │ ├── 003_add_cli.md
│ │ ├── 004_add_config.md
│ │ ├── 005_add_iterations.md
│ │ ├── 006_add_dynamic_plugins.md
│ │ └── plugins/ # Core plugins
│ └── tests/ # Test generation
├── output/ # Generated artifacts
│ ├── bootstrap/ # Bootstrap outputs
│ │ ├── bin/ # Bootstrap compiler
│ │ │ └── vibec.js
│ │ ├── bootstrap.js # Bootstrap script
│ │ └── test.sh # Bootstrap test script
│ ├── current/ # Latest merged runtime version
│ │ ├── bin/ # Current compiler
│ │ │ └── vibec.js
│ │ ├── bootstrap.js # Current bootstrap script
│ │ ├── test.js # Current test suite
│ │ └── test.sh # Current test script
│ └── stacks/ # Staged stack outputs
│ ├── core/ # Core stack stages
│ │ ├── 001_add_logging/
│ │ │ └── bin/
│ │ │ └── vibec.js
│ │ ├── 002_add_plugins/
│ │ │ └── bin/
│ │ │ └── vibec.js
│ │ ├── 003_add_cli/
│ │ │ └── bin/
│ │ │ └── vibec.js
│ │ ├── 004_add_config/
│ │ │ └── bin/
│ │ │ └── vibec.js
│ │ ├── 005_add_iterations/
│ │ │ └── bin/
│ │ │ └── vibec.js
│ │ ├── 006_add_dynamic_plugins/
│ │ │ └── bin/
│ │ │ └── vibec.js
│ └── tests/ # Test stack stages
│ ├── 001_basic_tests/
│ │ ├── test.js
│ │ └── test.sh
│ ├── 002_feature_tests/
│ │ └── test.js
│ ├── 003_cli_tests/
│ │ └── test.js
│ └── 004_config_tests/
│ └── test.js
├── vibec.json # Configuration file
└── package.json # Node dependencies
vibec employs a progressive bootstrapping process:
bin/vibec.js implementationvibec.js when generated in a stage, using the new version for subsequent stagesInstall globally:
npm install -g vibec
Or use via npx:
npx vibec --version
Set your LLM API key:
export VIBEC_API_KEY=your_api_key_here
Run with custom options:
npx vibec --stacks=core,tests --test-cmd="npm test" --retries=2 --iterations=3 --output=output
CLI options:
--workdir=<dir>: Working directory (default: .)--stacks=<stack1,stack2,...>: Stacks to process (default: core)--dry-run: Simulate without modifications (default: false)--start=<number>: Start with specific stage number (default: none)--end=<number>: End with specific stage number (default: none)--api-url=<url>: LLM API endpoint (default: https://openrouter.ai/api/v1)--api-model=<model>: LLM model (default: anthropic/claude-3.7-sonnet)--test-cmd=<command>: Test command to run (default: none)--retries=<number>: Retry attempts for API calls (≥ 0, default: 0)--iterations=<number>: Number of times to retry a stage on test failure (> 0, default: 2)--plugin-timeout=<ms>: Timeout for JS plugins in milliseconds (default: 5000)--output=<dir>: Output directory (default: output)--help: Display usage information--version: Show version (e.g., vibec v1.0.0)Configure via vibec.json:
{ "workdir": ".", "stacks": ["core", "tests"], "dryRun": false, "start": null, "end": null, "testCmd": "npm test", "retries": 2, "iterations": 3, "pluginTimeout": 5000, "apiUrl": "https://openrouter.ai/api/v1", "apiModel": "anthropic/claude-3.7-sonnet", "output": "output"
}
Option precedence: CLI > Environment Variables > vibec.json > Defaults
Validation:
retries: Must be non-negative (≥ 0)iterations: Must be positive (> 0)pluginTimeout: Must be positive (> 0)vibec.json triggers an error log and falls back to defaultsVIBEC_WORKDIR: Working directory path.VIBEC_STACKS: Comma-separated stacks (e.g., core,tests)VIBEC_DRY_RUN: true/false.VIBEC_START: Numeric stage value.VIBEC_END: Numeric stage value.VIBEC_OUTPUT: Output directory.VIBEC_TEST_CMD: Test commandVIBEC_RETRIES: Retry count for API callsVIBEC_ITERATIONS: Number of times to retry a stage on test failureVIBEC_PLUGIN_TIMEOUT: Plugin timeout (ms)VIBEC_API_URL: LLM API endpointVIBEC_API_KEY: LLM API key (recommended over config)VIBEC_API_MODEL: LLM modelVIBEC_DEBUG: Enable debug logging (1 to enable)Compatible with OpenAI-style APIs. Configure via VIBEC_API_URL and VIBEC_API_KEY.
Prompts use markdown:
# Component Name Description of the generation task. ## Context: file1.js, file2.js ## Output: path/to/output.js
## Context:: Reference files for context## Output:: Specify output file paths (multiple allowed).md)stacks/<stack>/plugins/.js)stacks/<stack>/plugins/{ config, // vibec.json contents stack, // Current stack name promptNumber, // Current prompt number promptContent, // Content of the current prompt workingDir, // Path to output/current testCmd, // Test command testResult // Test execution result
}
The iteration system allows automatic refinement of generated code:
iterations times (default: 2)This enables self-healing code generation that can fix test failures automatically.
stacks/core/005_new_feature.md)NNN_name.md naming convention## Output:Tests in stacks/tests/ generate:
test.sh: Validates vibec.js and runs test.jstest.js: Uses tape for unit tests (Node builtins only)This tutorial is based on a fully-functional example project available at: github.com/Strawberry-Computer/vibe-pong
mkdir pong-game && cd pong-game mkdir -p stacks/core/plugins output
Create a coding style plugin that forces the LLM to mark its reasoning in the generated code. This clever trick helps us trace which parts of the prompt produced specific code:
- Make it clear why any given code is written by adding a comment with relevant prompt snippet, like: ``` // PROMPT: <relevant prompt snippet> ``` - Don't use any other comments.
# Pong Game Initial Setup Generate a simple Pong game using HTML, CSS, and JavaScript: - Use `<canvas width="800" height="400" id="pongCanvas">` - Draw a white paddle (10px wide, 100px high) at the left side of the canvas, movable up/down with arrow keys. - Draw a white ball (10px radius) starting at canvas center, moving diagonally with constant speed. - Bounce the ball off the top and side walls; reset to center if it hits the bottom (misses paddle). - Detect paddle collision to bounce the ball back up. - Use requestAnimationFrame for smooth animation. ## Output: index.html ## Output: styles.css ## Output: game.js
{ "stacks": ["core"], "output": "output"
}
export VIBEC_API_KEY=your_api_key_here npx vibec
cd output/current python3 -m http.server 8000
Visit http://localhost:8000.
# Add Scoring to Pong Enhance the Pong game by adding score display ## Context: index.html, game.js ## Output: index.html ## Output: game.js
Run npx vibec again to update the game.
# Add AI Player to Pong Add the computer player which moves the paddle automatically. Make sure that at this point there are 2 paddles: one for the user and one for the computer. ## Context: index.html, game.js ## Output: game.js
Run npx vibec one more time to complete the game with an AI opponent.
Use VIBEC_DEBUG=1 npx vibec for detailed logs.
VIBEC_API_KEY## Output: in promptsnpx vibec or install globally--plugin-timeout if needed
