preparelisted

# Pre-Implementation Analysis: $ARGUMENTS Analyze BEFORE implementing. Do NOT write production code — only analyze, plan, and (on confirmation) save decomposed plans. Previous step: `/analyst` (clarify requirements). Next step: `/implement` (one invocation per subtask file). **Input is one card / one coherent change — check the tier before Phase 0.** A whole-surface audit or a program that plausibly yields 20+ cards ("review every screen", "audit this whole area") is **tier 1**, and this skill is tier 2: run **`/prompt-master`**'s program pass first (breadth sweep → `00-report.md` + `evidence/` + `cards/NN-<slug>.md` + `RUN-ORDER.md`, stopping before any implementation), then come back here **one card at a time**. Decomposing a surface directly into waves skips the discovery that decides which cards are worth cutting at all. > **Risky work → suggest plan mode.** For a change the user flags as risky (production data, > irreversible migrations, wide blast radius), recommend running the `/implement` session > under Claude Code's built-in **plan mode** (plan approval before any edit) — it is the > harness-level twin of this skill's plan-first pipeline: the kit plans the work, plan mode > gates the execution. `/rewind` checkpoints add per-edit undo on top. --- ## Deliver, don't halt — ask only when the choice is material (governing rule) This skill's job is to hand back the **finished analysis + plan**. An undertold detail must never end the run at "tell me (a)/(b)/(c)". Sor