sdd-plan-writerlisted

# SDD Plan Writer ## Purpose An AI coding agent stops when the work _looks_ done. Without a check the agent itself can run, "looks done" is the only signal available, and every silent mistake waits for a human to notice — this is Anthropic's own diagnosis of why plans fail in practice, not a hypothetical concern (see Anthropic's Claude Code best-practices documentation). A spec that says "add validation to the signup form" leaves the agent free to satisfy that sentence in a dozen mutually-inconsistent ways; a spec that names the exact file, the exact trigger/response behavior, and the exact command that proves it worked leaves no such freedom. This skill exists to produce the second kind of spec, not the first — every task must carry its own path, acceptance criterion, and verification command, or it does not go into the plan yet. ## When to use **Precondition — the input must already be a finite, confirmed list of work items.** Each item must already be understood well enough to name at least one concrete file it touches. Typical inputs: the Core-gap rows of a closed `axis-closure-matrix` report, an already-agreed feature description, a named bug with a known repro. If any item still requires "figure out what's actually wrong first," send that item through investigation (a general-purpose research subagent, or your own direct reading) BEFORE this skill runs — do not fold an open investigation into a plan task; a plan task is for **known, scoped** work, not for discovery.