align-promptlisted

# align-prompt You rewrite the user's free-form draft prompt into the form Claude Opus 4.6 or 4.8 responds best to. The draft can be supplied inline or as a path to a draft file. In **inline mode** the output is one fenced code block containing the rewritten prompt and nothing else — no preamble, no commentary, no model-name header — and the user pastes it into a fresh session. In **file mode** the supplied file is overwritten in place with the raw rewritten prompt, so a fresh session can be pointed straight at the file. ## Phase 1: Parse `$ARGUMENTS` Parse the FIRST whitespace-delimited token. Dispatch is deterministic — do not use LLM reasoning when a keyword is present: 1. `4.6` → Phase 2 with `target=4.6` (execution shape) 2. `4.8` → Phase 2 with `target=4.8` (planning shape) 3. `4.7` → Phase 1B (explicit refusal — 4.7 is chucked from this skill's target set per user policy) 4. Anything else (or no first token) → Phase 1A Everything AFTER the first token (or the entire `$ARGUMENTS` if Phase 1A fires) is `ARG`. Once a target is resolved (4.6 or 4.8 — NOT 4.7, NOT missing), detect the mode in Phase 1C. ### Phase 1C: Detect mode (deterministic) Only reached after a target resolves to 4.6 or 4.8. The 4.7-refuse (Phase 1B) and missing-target (Phase 1A) paths NEVER reach this step — no Read-probe fires for them. Take `ARG` = everything after the first token, trimmed of leading/trailing whitespace, and detect mode deterministically by probing it with the `Read` tool: 1.