Large language models (LLMs) have made incredible strides, but their monolithic nature presents challenges for building reliable, modular systems. Think of it like trying to build a car in one giant, inseparable piece. How do you fix a problem with the engine without affecting the wheels? This is where specifications come in. Just as blueprints guide the construction of complex machinery, clear specifications are the missing link to making LLM development a true engineering discipline. This concept, explored in a recent research paper, argues that the key to unlocking the next level of AI lies in precise descriptions of what each AI component should do and how to verify its actions. These descriptions, known as 'specifications,' are crucial for five key properties: verifiability, debuggability, modularity, reusability, and automated decision-making. Currently, LLMs rely on natural language prompts, which are inherently ambiguous. This ambiguity makes it hard to determine if an LLM is functioning correctly and even harder to fix issues. It's like giving someone vague cooking instructions – you might get something edible, but it probably won't be what you envisioned. The research paper suggests various techniques to address this ambiguity, including iterative disambiguation (where LLMs ask clarifying questions), domain-specific rules, and structured outputs (using formats like JSON). Imagine being able to build complex AI systems by combining pre-built, reliable LLM components, much like assembling LEGO bricks. With clear specifications, developers could easily swap components, debug individual parts, and build far more sophisticated AI systems. This shift towards modularity is crucial for cost-effectiveness, faster improvements, and easier integration with existing software. The future of AI doesn’t just lie in bigger models, but in smarter, more reliable systems built on a solid foundation of specifications.