The Kotlin Creator’s Next Act: CodeSpeak Wants to Replace Your Code with Specs

What happens when the person who designed one of the most popular programming languages decides that traditional code is the problem? Andrey Breslav — the mind behind Kotlin — is betting that in the age of LLMs, engineers should stop maintaining code entirely and start maintaining specifications instead. His new project, CodeSpeak, is a programming language that compiles plain-English specs into Python, Go, JavaScript, and TypeScript, promising to shrink codebases by 5–10x.

The project hit Hacker News on March 12, 2026, pulling in 206 points and 169 comments. Gergely Orosz dedicated a full episode of The Pragmatic Engineer podcast to it — a nearly two-hour conversation. Developers like Pamela Fox weighed in on Twitter/X. The discussion is heated, and it touches a nerve: what should programming look like when AI can handle the boring parts?

What CodeSpeak Actually Does

CodeSpeak is not vibe coding. It’s not a chatbot wrapper. And it’s not a prompt-engineering framework.

The core idea is deceptively simple: you write a specification file — structured English with modules and reusable components — describing what your program should do. CodeSpeak then compiles that spec into real production code in your target language. Think of it as a higher-level programming language where the “syntax” is constrained English and the “compiler” is an LLM.

The workflow looks like this:

Write a spec describing your module’s behavior, inputs, outputs, and edge cases
Run codespeak build to generate production code
Tests run automatically to verify the generated code matches the spec
When you need changes, edit the spec — not the generated code

Version 0.3.1 introduced a takeover command that goes the other direction: point it at existing source code, and it extracts a spec you can then maintain going forward. This is how CodeSpeak envisions migration for existing projects — not a rewrite-from-scratch, but a gradual conversion.

Under the hood, CodeSpeak uses Anthropic’s Claude API with a bring-your-own-key model. You install it via uv tool install codespeak-cli and can start converting files immediately.

The Numbers: How Much Code Actually Disappears?

CodeSpeak’s website showcases several real-world conversions, and the reduction ratios are striking:

WebVTT subtitle parser: 255 lines of code → 38 lines of spec (6.7x reduction)
Italian SSN generator: 165 lines → 21 lines (7.9x reduction)
HTML/XML encoding utility: 826 lines → 141 lines (5.9x reduction)
EML email converter: 139 lines → 14 lines (9.9x reduction)

The team also demonstrated fixing a real GitHub issue in Microsoft’s MarkItDown project: 23 spec lines generated approximately 221 lines of production code — a 10x expansion ratio. The generated code included helper methods, attachment handling, and automated test vectors, all passing the test suite.

These are utilities and converters, though — the kind of code where behavior is well-defined and deterministic. The question is whether these ratios hold for messier, more stateful applications.

“Maintain Specs, Not Code” — But Can It Scale?

CodeSpeak’s tagline is “maintain specs, not code,” and Breslav’s argument is philosophical as much as practical. A huge percentage of application code, he argues, consists of tedious details that are obvious to any experienced developer. Now that LLMs can also recognize these obvious patterns, why should humans write them out by hand?

The language supports mixed-mode development — you can have some files managed by CodeSpeak specs and others written traditionally in the same project. This is a pragmatic concession that not everything fits the spec-to-code model, at least not yet.

Breslav positioned this clearly in The Pragmatic Engineer interview: CodeSpeak is designed to preserve “the essence of software engineering” — the business logic, the architecture decisions, the tricky edge cases — while offloading everything else. His framing: “Everything else, the machine knows as well.”

But the developer community is split. Pamela Fox, a well-known developer advocate, noted on Twitter/X that she can see CodeSpeak working well for utilities and pure functions but has “a hard time imagining it for full-stack apps.” This is the central tension. Spec-driven development works beautifully when inputs, outputs, and behaviors are cleanly defined. It gets murkier when you’re dealing with state management, UI interactions, database migrations, and the messy reality of production systems.

CodeSpeak vs. the Current AI Coding Landscape

CodeSpeak enters a crowded field, but it’s playing a different game than most.

Cursor and Claude Code are AI-assisted coding tools — they help you write and edit traditional code faster. You’re still maintaining Python, JavaScript, or whatever language you’re working in. The AI is a copilot, but the artifact is still code.

Vibe coding tools (like Replit Agent or Bolt) let non-engineers describe what they want and get a working app. They’re optimized for speed and accessibility, not for long-term maintainability by engineering teams.

CodeSpeak sits in a third category: it’s a new abstraction layer. You maintain specs, the compiler produces code, and you never touch the generated output directly. It’s closer to how SQL works — you declare what you want, the engine figures out how to do it. Except the “engine” here is an LLM, which introduces its own set of reliability concerns.

The key differentiator is that CodeSpeak is explicitly designed for professional engineering teams working on long-term projects, not for prototyping or one-off scripts. Whether that positioning holds up depends entirely on how deterministic and reliable the spec-to-code compilation can be. If the same spec produces different code on different runs, that’s a dealbreaker for production use. CodeSpeak’s test-suite integration is clearly designed to address this, but it’s early days.

Why Breslav’s Track Record Matters

It would be easy to dismiss CodeSpeak as another AI hype project. The reason people are paying attention is Breslav’s resume. He didn’t just “contribute to” Kotlin — he created it. He led its design at JetBrains through multiple releases, navigated the treacherous challenge of Java interoperability, and saw it through to becoming Google’s preferred language for Android development.

Building a successful programming language is one of the hardest things in software engineering. The fact that Breslav has done it before — and done it well — gives CodeSpeak a credibility floor that most new language projects never achieve. He was also invited to give a talk at Cambridge’s Department of Computer Science on the topic of transitioning from Kotlin to CodeSpeak, further signaling that the academic community is watching.

That said, credibility isn’t a guarantee. Kotlin succeeded partly because of JetBrains’ tooling muscle and Google’s endorsement. CodeSpeak doesn’t have those institutional tailwinds — yet.

FAQ

Is CodeSpeak free to use?
CodeSpeak itself is available as a CLI tool you can install via pip/uv. However, it requires an Anthropic API key to function, so your costs depend on your Claude API usage. There’s no separate subscription fee for CodeSpeak at this time — it’s a bring-your-own-key model.

What languages does CodeSpeak compile to?
Currently Python, Go, JavaScript, and TypeScript. The team has demonstrated conversions with real-world Python projects like yt-dlp, Faker, BeautifulSoup4, and MarkItDown.

How does CodeSpeak compare to GitHub Copilot or Cursor?
They solve different problems. Copilot and Cursor help you write traditional code faster — you’re still the one maintaining that code. CodeSpeak replaces the code entirely with a higher-level spec. Think of it as the difference between a faster typewriter and a dictation system.

Can I use CodeSpeak on an existing project?
Yes. The takeover command in version 0.3.1 lets you convert existing source files into CodeSpeak specs. Mixed-mode development means you can convert files gradually rather than migrating everything at once.

What’s the current status of CodeSpeak?
It’s in alpha (version 0.3.1 as of early 2026). The team is actively hiring AI researchers and engineers. It’s functional and usable for real projects, but expect rough edges and rapid changes.

Top AI Product