CodeSpeak: Write Specs, Not Code
What if you could replace hundreds of lines of code with a 20-line spec and let an LLM handle the rest? That is exactly what CodeSpeak promises. Created by the inventor of Kotlin, CodeSpeak is a next-generation programming language that uses LLMs to generate production-grade code from human-readable specifications.
What is CodeSpeak?
CodeSpeak is a formal language for communicating with LLMs. Instead of writing implementation code, you write .cs.md specification files that describe what you want. The codespeak build command then generates working code that passes your test suite.
The project claims 5-10x codebase reductions in real-world case studies. For example, WebVTT subtitle support for yt-dlp went from 255 lines of code to just 38 lines of spec, a 6.7x shrink factor, while all tests continued passing.
Key Features
- Mixed mode adoption: Use CodeSpeak for new features while keeping existing code untouched
- BYOK (Bring Your Own Key): Works with your Anthropic API key
- Test-driven generation: Specs are validated against your existing test suite
- Version control friendly: Specs are markdown files that diff cleanly
- Deterministic builds: Same spec produces consistent output
Installation
Getting started takes two commands:
uv tool install codespeak-cli
codespeak login
Initialize a mixed-mode project to add CodeSpeak features alongside existing code:
codespeak init --mixed
Usage
Create a spec file describing your feature:
# EmlConverter
Converts RFC 5322 email files (.eml) to Markdown.
## Accepts
`.eml` extension or `message/rfc822` MIME type.
## Output Structure
1. **Headers section**: From, To, Subject, Date
2. **Body**: plain text preferred
3. **Attachments section**: list with filename and size
Register the spec and build:
codespeak build
CodeSpeak generates the implementation code, wires it into your project, and verifies tests pass.
Operational Tips
Start with a small, isolated feature to evaluate the approach. The mixed-mode workflow makes adoption low-risk since existing code stays untouched. For complex builds, set Claude Opus 4.6 as your model for better results.
Conclusion
CodeSpeak represents a shift toward intent-driven development. Instead of telling the computer how to do something, you describe what you want. For DevOps teams drowning in boilerplate or maintaining large codebases, this could dramatically reduce cognitive load.
Check out the CodeSpeak documentation to experiment with spec-driven development.
For AI-powered operations automation, Akmatori helps SRE teams build intelligent agents that handle incidents, runbooks, and infrastructure tasks. Powered by Gcore infrastructure for global reliability.