An implementation plan is the bridge between a user story (what to build) and the code (what was built). It is written before the Developer persona session begins, and it breaks the story into a concrete sequence: which files to modify, which interfaces to implement, which tests to write, and in what order. The plan should be specific enough that the AI can follow it without improvising, and concise enough that a human can review it in under ten minutes.
The implementation plan is not a design document — the Architect's ADRs handle that. It is an execution script for a bounded session. It references the relevant ADRs by decision ID and the acceptance criteria by story ID. If the plan cannot reference either, the story is not ready to implement.
Plans also serve as a retrospective quality tool. If the AI followed the plan and the output is correct, the plan was good. If the AI followed the plan and the output was wrong, the plan was incomplete — a signal to improve the template for future stories. If the AI deviated from the plan, the deviation needs a human decision before merge. Undocumented deviations from implementation plans are where architecture debt accumulates.