What if the biggest productivity problem with AI coding agents isn't the model, but how you structure the conversation with it?
A workflow called Spec-Driven Development (SDD), built around Claudee Code](/tools/claude-code/), argues that most developers are setting AI agents up to fail by handing them large, loosely-defined tasks and hoping the model figures out the rest. The approach offers a more structured alternative built on two ideas: decompose before implementing, and clear the model's working memory between every phase.
How the Decomposition Actually Works
SDD breaks coding work into two distinct layers. The first layer is specification generation - before writing any code, you produce a series of documents in sequence: a requirements document, a code analysis of the existing codebase, and a design document outlining the approach. Each document is a separate step, and the AI's context window (the amount of prior conversation the model can "see" at once - think of it like short-term memory) is cleared between each one.
The second layer is implementation. Once specs exist, the overall task is split into small, discrete subtasks. Each subtask gets its own agent session with the relevant spec attached. Context is cleared again after each subtask completes.
The result is a model that's never trying to hold the whole problem in view at once. Each session has a narrow, well-defined job.
The Case for Forgetting on Purpose
Clearing context between steps sounds backward - why would you want the AI to "forget" what it just did? The argument is that large context windows produce worse results on focused tasks, not better. When the model can see hundreds of prior messages, earlier decisions and irrelevant details start influencing current output in subtle ways. A clean slate means the model works only from information that's actually relevant to this step.
There's also a direct cost benefit. Longer context means more tokens processed per request, which means higher API bills. SDD practitioners report meaningfully lower per-cycle costs compared to single long-context sessions on the same problem.
This workflow makes the most sense for developers using Claude Code on tasks involving multiple files, cross-cutting dependencies, or real architectural decisions. For a quick one-file fix, it's overhead you don't need. But if you've repeatedly hit the ceiling on what AI coding agents reliably produce - getting plausible-looking code that doesn't actually fit together - SDD addresses that as a structural problem in how you're using the tool, not a quality problem with the model itself.