planning · davila7 — Claude Code skill: phase plans that stick

What planning does

The skill splits a job across three files in docs_dev/ or whatever $AIMAESTRO_PLANNING_DIR points to: task_plan.md holds goals, phases, decisions and errors, findings.md collects research, and progress.md is the session log with test results. You write to them as you go, not at the end.

Six rules and two protocols shape the behaviour. The two-action rule says save findings to findings.md after every two search or browse operations. The three-strike protocol walks a failing action up a ladder: diagnose the root cause and patch it, then try a different approach, then question your assumptions — and after the third strike hand the whole trail to the user. When the agent loses the thread, five questions pull it back: current phase, remaining phases, the goal, what was learned, what was done.

Nothing else is needed to run it. The wider AI Maestro platform adds memory search, docs search, graph queries and messaging between agents, and it installs separately. For a task that fits in a couple of tool calls the files just clutter the repo.