planner

Define the phase goal, then decompose it into plans of 2-3 tasks each. Every task needs exact file paths, specific actions with pitfalls to avoid, an executable verify command, and measurable done criteria. Group independent plans into the same execution wave.

Goal-backward planning starts from the done state and asks what must be true for the goal to be achieved, rather than listing tasks forward. It produces must-have truths and artifacts that every task must satisfy, ensuring plans stay aligned with the actual objective.

Each plan should contain 2-3 tasks maximum. Plans exceeding 3 tasks, touching more than 5 files per task, or crossing multiple subsystems should be split into separate plans to keep AI execution quality high and context usage under roughly 50 percent.

No. Plans assigned to the same execution wave must not modify the same files. If two plans need the same file, move one to a later wave or split the file into separate modules to avoid conflicts during parallel execution.

Use TDD plans for complex business logic with edge cases, financial calculations, state machines, data transformation pipelines, and input validation rules. A TDD plan follows Red, Green, Refactor phases where failing tests are written before implementation.

Large plans push AI context usage past 50-70 percent, where output quality degrades into rushed, minimal work. Aggressive atomicity with more plans of smaller scope keeps each execution within peak quality range and makes verification tractable.