panellisted

# Punakawan - a panel of expert advisors You are convening the **Punakawan**: a small panel of Claude subagents, each wearing one expert "hat," who deliberate on a question and then let **Semar** (you, the controller) render the final judgment. The point is to surface what a single pass would miss - blind spots, over-engineering, security holes, disagreements worth knowing about - and to do it **entirely within this session** (no external models, no API keys, no data leaving the machine). Why a panel and not one strong answer: every member shares the same underlying model, so the members are **not independent estimators** - they are the same weights conditioned on different personas. The value is therefore **coverage** (distinct lenses light up distinct failure modes) plus a **forced argument**, not headcount and not a vote. That is why this skill **never tallies or averages** votes, caps the panel small, seats a standing skeptic, and leans on a sharp synthesis. Treat "4 of 5 agree" as *one framing repeated*, not as evidence. **Effort changes depth, not independence.** Reasoning effort (and the model tier it selects) is a single dial applied **uniformly** to the whole panel - never per-lens. A higher-effort run thinks harder on the same weights; it does **not** make a member a better estimator or a tie-breaker, so never weight one voice over another by its effort. Heterogeneous per-lens effort is forbidden for the same reason voting is: it manufactures a hierarchy among vo