A practical workflow for LLM-assisted development

The process of working with language models (LLMs) can feel magical, but also challenging when they fail. After months of daily use, the author developed intuition for when LLMs produce useful code and when they fail. To ensure reliable results, it's important to limit scope and provide enough scaffolding. LLMs are powerful tools that can help build projects on a scale that would be difficult to attempt manually.

They generate a lot of code out of the gate, so the focus shifts to refining the code to fit specific needs. The agentic loop can be viewed as a genetic algorithm, where the model outputs something roughly correct, receives feedback, and iterates on the code until it converges on a solution. LLMs are trained on massive amounts of public code, making them excellent at completing typical tasks, such as writing service endpoints or building a UI.

They can also perform explorative work, like identifying call graphs and tracing steps to understand service endpoint implementation. LLMs excel at handling language-specific syntax, allowing developers to work in unfamiliar languages more efficiently. While LLMs don't replace human skills, they can amplify what an expert can achieve by allowing faster work and focusing on the big picture.

The main challenges with LLMs include dealing with context and creativity. Developers must provide explicit logic, such as specifying the runtime environment and authoritative sources. Additionally, agents may generate solutions with structural issues, such as incorrect path handling or inefficiencies. To use LLMs effectively, one should have a solid understanding of the project before starting and spell out constraints to guide the agent towards the desired solution.