CodeSpeak, from Kotlin creator Andrey Breslav, shifts development from code to specifications

A team led by Andrey Breslav, the creator of Kotlin, has unveiled a public alpha of CodeSpeak, a tool that moves the center of development from writing code to describing a system's architecture and behavior. The developer writes specifications in English, and LLMs handle generating, testing and refactoring the executable code.

How CodeSpeak works

The core idea of CodeSpeak is to make the specification the main artifact of a project and turn code into a derived layer. Instead of manually writing dozens of files, describing interfaces, keeping the module structure in their heads and tracking syntax details, the engineer sets requirements in natural English. The model then generates executable code, fills in missing parts, updates tests and helps with refactoring when the product or architecture changes.

In this approach, the emphasis shifts from lines of code to the intent behind the system: what the service must do, what constraints apply, and how components should interact. For teams, it looks like an attempt to raise the level of abstraction by one more step. Developers once moved from machine instructions to high-level languages, then to frameworks and infrastructure as code. The next step, in this view, is development through formal text specifications.

What the platform promises

According to the team, with full adoption CodeSpeak can reduce a project's codebase five to ten times. This does not mean code disappears entirely. Instead, a significant share of boilerplate, glue and repetitive work moves to automatic generation. That leaves people with less to maintain by hand, while key decisions concentrate in specifications, which are easier to read, discuss and review.

At launch, the platform stands out for several practical reasons:

• the developer describes system behavior in English rather than assembling everything by hand from syntactic constructs
• LLMs take part not only in generation but also in testing and refactoring code
• the tool is intended not only for new projects but also for integration into existing complex systems
• among the stated scenarios is work with a Python codebase, which lowers the barrier to a real pilot

This detail matters because many AI tools look good in demos but break when they meet real legacy code, complex project structure and established engineering practice. Here the team speaks right away about embedding into existing Python projects. That means it is betting not only on greenfield development but also on a more painful scenario for the market: gradually modernizing what already works.

Where it may be useful

judging by its positioning, CodeSpeak may fit best where a large codebase is expensive to maintain and onboarding new developers takes a long time. Examples include internal platforms, product backend services, integration layers and corporate tools with many rules and checks. In such projects, the value often lies not in elegant syntax but in transferring business logic into a working system quickly and accurately, without constantly rewriting repetitive pieces by hand.

But this approach has a hard requirement: specifications must be precise. If a team describes system behavior vaguely, the model will fill the gaps with its own assumptions. So CodeSpeak is unlikely to remove the need for strong engineers. Rather, it changes their role: less manual assembly, more architectural thinking, defining constraints and checking that the generated code actually matches the intent. For a public alpha this is especially important: human oversight remains mandatory here.

What it means

CodeSpeak shows how AI tools are moving from the "suggest a snippet of code" format toward a model in which the specification becomes the source of truth and the LLM becomes the execution layer of development. If this approach takes hold in real Python teams, the market will get not just another AI assistant but a new level of abstraction for engineering work.