Pools of Extraction: How I Hack on Software Projects with LLMs (2025) (blog.almaer.com)

In a recent reflection on software development practices, a developer highlighted the concept of "pools of extraction," demonstrating how specifications and context can enhance collaboration and efficiency in building projects. By using AI, particularly large language models (LLMs), to create and refine specifications, the developer emphasizes the importance of clear communication and shared sources of truth, akin to the iterative design process between an architect and a client. This approach enables teams to better align on project goals, reducing misinterpretations commonly associated with informal discussions, and allows specifications to serve as a reliable foundation for future work. The significance for the AI/ML community lies in the nuanced understanding of how both LLMs and human cognition operate within these "pools." LLMs possess a "latent pool" of knowledge from training data and a "contextual pool" that relies on the specific input provided during a session. The author argues that by narrowing focus—such as through small, experimental projects or by extracting relevant pieces of code—developers can maximize the efficiency and relevance of AI outputs. This reinforces the crucial role of context alongside the technical power of LLMs, indicating that both effective memory retrieval in humans and context awareness in AI are key to improving output quality and facilitating creative problem-solving.