Summary
Today’s best AI content shares whole projects of skills, plugins of skills and agents, and spec-driven flows. I use them every day. But they are shown on small projects that fit in the assistant’s head, and none of them takes into account the project that grows feature after feature. That is where the assistant fades: the model has not changed, what it has to hold has grown.
Partition is the method I packaged for that project, as agent skills for Claude Code and GitHub Copilot that run beside the tools you already use. Boundaries come from a source and you decide the split. The spec moves first and you validate every rule. The state lives in the plan, not the chat. And the budget is measured: what each task loads before reading code, against a threshold, with a context split once when it grows past it.
The proof is public: partition-commerce, the outcome of the 20-step guide, grew from an empty skeleton to two deployables and five bounded contexts, and its heaviest context loads a quarter of the threshold. Across the many projects the method ran on before its publication, the mean gains are 56% on delivery time and 45% on AI credits.
Key ideas
- The model has not changed; what it has to hold has grown, and every request pays for it.
- Skill packs, plugins and spec-driven flows are shown on small projects; Partition completes them for the project that grows.
- Boundaries from a source, the spec moves first, the state lives in the plan, the budget is measured.
- partition-commerce, the outcome of the guide, keeps its heaviest bounded context at 25% of the threshold.
- Across the projects that use it, the mean gains are 56% on delivery time and 45% on AI credits.
Why I wrote this
The tools shared today are shown on small projects and leave out the project that grows feature after feature; this article presents the method for that project.