Create Plan Artifact

Generate .agent/{branch-name}/plan.md — the planner→executor handoff document — alongside checks.yaml, the executable acceptance checks derived from the Acceptance Criteria.

A new Claude session MUST be able to execute from plan.md alone without the original conversation.

plan.md is a handoff artifact, not an exhaustive knowledge base. Keep it lean (see the Core/Extended tiering in Step 2). checks.yaml is the living contract — it is re-derived on every iteration (statuses reset to pending) and self-validates against reality on every executor loop, so it does not go stale the way prose does. When something is unclear, prefer asking the still-running planner agent (or the user) over bloating the document to answer every future question.

Snapshot mode (opt-in)

By default this skill writes only plan.md (overwritten in place each iteration; the version: frontmatter is the iteration counter). .agent/ is gitignored per-developer scratch, so immutable plan.vN.md snapshot chains are not written by default — they are rarely re-read and cost a redundant write per iteration, with no measured effect on task success (research §5.4).

Opt in to snapshots with Skill("aw-create-plan", "snapshot") — then the skill also writes an immutable plan.v{N}.md alongside plan.md (plan.v1.md, plan.v2.md, … as history) for a durable audit trail of the plan's evolution. Everything below marked (snapshot mode) applies only then.

How the mode arg is read (disambiguation)

The arg slot is overloaded — some callers pass a short mode flag, others a whole plan body. Resolve it before anything else: snapshot mode is on iff the trimmed arg is exactly snapshot or --snapshot (case-insensitive, a single bare token). Every other arg — empty, or a multi-line plan.md body — is default mode (no snapshot). In particular, orchestrators like /fix-bug pass the whole plan body here (Skill("aw-create-plan", "<full plan.md body>")), which is default mode by design — the fast lane opts out of snapshots anyway.