Amazon’s Kiro uses the Easy Approach to Requirements Syntax (EARS) in its specification-driven development workflow. It popularizes a practice: turn a vague prompt or context into a specification (spec) in a clear format, so both humans and agents can better understand.
The EARS was developed by Alistair Mavin and colleagues at Rolls-Royce and has become a widely adopted notation for writing clear, testable requirements. Its original paper won the 10-year most influential industry paper award in the IEEE International Requirements Engineering Conference 2019.
However, the notation was originally designed for high-level stakeholder requirements—not for the full spectrum of software specifications (specs) that modern development demands. Working with EARS in spec practice revealed friction points.
Therefore, we propose GEARS, generalized EARS or Generalized Expression for AI-Ready Specs. It extends EARS for the age of AI coding by preserving what made EARS successful while adapting the syntax for broader use in specs.
What Are Specs
Specs are natural-language descriptions of system requirements and behaviors. Specs should be versioned and maintain "eventual consistency" with code so that they become the primary source of truth for AI to understand the system. In a broader sense, specs can also describe the decisions, plans, and changes made to a system, though GEARS doesn’t cover them as they are auxiliary and less frequently cited in prompts or contexts for AI.
Specs are the new “source code” for human developers. We believe specs are essential to AI-era software development for two reasons:
- Without clear specs, misunderstandings among humans and LLMs constantly exist. Specs are the natural-language media for human and AI developers to communicate.
- There are always divisions of work in software even with the power of AI. Clear specs facilitate humans and LLMs to understand and reuse system components.
Specs are iterative. It is not a return to the waterfall model. Specs grow from scratch along with the code iteration by iteration.
Specs can be AI-generated. It doesn’t mean humans have to write every single line of specs. Usually human developers start with high-level and sometimes vague intent and discuss with AI. Then AI can help specify, propose considerations, and fill in details. Finally, AI can write the specs in the right format.
Why Use GEARS
AI needs consistency. LLMs perform better with predictable structure. A spec format acts as a protocol between human intent and AI execution. Without it, misunderstandings easily propagate across iterations and create volatility and risk.
Humans need it too. The bottleneck in AI coding is rarely the LLM’s capability—it’s our ability to express what we want. A constrained syntax leads to clarity. You don’t want to write a free-form prose; the structure helps us get precision.