Summary
The name holds two meanings: the score musicians play, corrected first when the music must change, and the split of a system into parts that can each be held in full. Partition is the method I apply with AI coding assistants, and it rests on both.
Today’s content shares projects of skills, plugins and spec-driven flows, and I use them every day, but they are shown on small projects. None takes into account the project that grows feature after feature, where the assistant loses accuracy. The Context Budget article stated that problem; Partition makes the method installable, as skills for Claude Code and GitHub Copilot, with an always-on instruction: the durable spec is the truth, the state lives in the plan, bounded contexts come from a source, and the budget is measured.
Its public project, partition-commerce, is the outcome of the 20-step guide: a split, three absorptions, a removal, an inherited spec and a promotion later, the heaviest context loads 25% of the threshold. Across the many projects it ran on before its publication, the mean gains are 56% on delivery time and 45% on AI credits.
Key ideas
- A score is corrected before the music changes; a spec moves before the code does.
- The tools shared today stop where the project starts to grow; Partition adds to them rather than replacing them.
- The durable spec is the truth, the state lives in the plan, bounded contexts come from a source, the budget is measured.
- partition-commerce shows a project that lived, split and grew, with its heaviest context at 25% of the threshold.
- Mean gains of 56% on delivery time and 45% on AI credits across the projects that use it.
Why I wrote this
The Context Budget article stated why a growing project costs every request; this one presents the method, packaged so that it no longer depends on memory or on the discipline of each session.