second-opinionlisted

# Second Opinion Stress-tests a verdict and re-emits it with per-claim grouping. Companion to `human-voice`, which produces the verdict this skill validates. ## Invariant **ALWAYS** confirm a subagent-spawning tool exists before starting, and report counts that equal what actually ran. **NEVER** run unbidden, and never emit this skill's output format without having dispatched the work it describes. ## Step 0 — Preconditions Both must hold before anything else happens. 1. **The user asked.** Either they invoked this skill, or `human-voice` offered it and they accepted. **Never auto-run.** Offering means one line naming the validation and stopping; the user's next message is the only trigger. 2. **A subagent-spawning tool is actually available.** Name the tool you will call. If you cannot name it, you are in the ungated case: > State that subagent validation is unavailable here, give the verdict's > uncertainty plainly, and offer search-based verification instead. **In the ungated case this skill's output format is forbidden** — no grouped Verified/Flagged/Conflict block, no `Δ Validation` block, no persona attributions, no `❗`. Respond in `human-voice` default mode. Producing the shape of a validation that did not run is the single worst failure this skill can have. Never run for quick facts or low-stakes picks. Cost consciousness is a design constraint, not a preference. ## Step 1 — Extract claims Decompose the verdict into load-beari