blueprint

To generate a multi-session engineering plan, you need a structured construction plan that decomposes large features into dependency-aware, PR-sized tasks. This provides self-contained context briefs so different agents can execute work cold across multiple sessions without losing context.

The best way to break down a large refactoring project is to decompose the objective into PR-sized tasks with clear dependency graphs. This approach ensures continuity and allows parallel execution detection to manage independent work streams efficiently.

Adversarial review improves project planning quality by delegating plan validation to a sub-agent that checks against a checklist of anti-patterns. This automated review step catches structural flaws and ensures the generated construction plan is robust before execution begins.

You need a Claude Code environment to run the planning agent, while Git and GitHub CLI are optional dependencies for automated branch and PR workflow management. The core planning functionality works without them, but they enable automated branch creation.

Yes, this planning approach is specifically designed for cold-start execution. It generates self-contained context briefs for every step, allowing entirely different agents to pick up and execute specific tasks without requiring prior knowledge of the overall project.

You should avoid using automated multi-session construction plans for simple, single-session tasks that do not require coordination across multiple agents. The overhead of generating dependency graphs and adversarial reviews is only justified for complex, multi-phase engineering projects.