plan

Imported from Xaxis/preflight (skills/plan/SKILL.md). Install upstream with npx skills add Xaxis/preflight --skill plan. Copyright stays with the author.

Produce a phased plan for: $ARGUMENTS

No edits until the human approves. Planning is the cheapest place to be wrong — a bad assumption costs one paragraph here and three phases of rework later.

Steps

1. Enter plan mode (EnterPlanMode) if you aren't already in it. It makes "no edits yet" a guarantee rather than an intention.

2. Know the ground first. If the target area is unfamiliar, run /preflight:research before planning. Planning against a guessed codebase produces a plan that dies on contact. Skip only if you already have the map.

3. Delegate the plan to the planner subagent, passing the goal, the research map, and any constraints the human gave. Its plan is a proposal to you, not a verdict — check it against what you know before endorsing it.

4. Resolve the real forks. If a decision genuinely changes what gets built and you can't settle it from the code, ask (AskUserQuestion). Don't ask what a sensible default already answers, and don't ask for approval — that's what step 5 is for.

5. Present for approval (ExitPlanMode). One recommended approach with phases, not a survey.

What a phase must have

• Goal — one sentence.
• Changes — specific files and what happens in each.
• Checkpoint — the exact command and the observable result that proves it worked.

A checkpoint is a command with an expected result: npm test -- auth.spec.ts passes. "Tests pass" isn't one. "It compiles" isn't one — compiling is not evidence of behavior. If a phase has no honest checkpoint, it's drawn wrong: reslice it into something that runs end-to-end.

Prefer thin vertical slices that work over horizontal layers that can't be verified until three phases later.

Execute

After approval, work the phases in order and stop at each checkpoint. Delegate checkpoint runs to test-runner so test output stays out of this conversation.

If a phase reveals the plan was wrong, stop and say so. Re-plan the remainder rather than improvising past a broken assumption — the plan was approved on premises that no longer hold, so silently continuing means shipping something the human never agreed to.