HerbertGao/herbertgao-skills

Multi-perspective council for one open decision, including architecture selection when a draft already exists and the goal is to choose rather than edit. It records four named first positions before synthesis and compares challengers under a neutral brief; missing isolation lowers assurance instead of blocking the debate. Canonical records may reach CONVERGED; weaker hosts return unaudited, non-authorizing ADVISORY. Use review-loop to find and fix artifact defects to APPROVE. Requires ~/.agency-agents.

Portable visual explainer that turns a complex concept or repository flow into one grounded, self-contained HTML file with large diagrams and few words. Adapted from the Anthropic community ELI5 skill.

Orchestrated OpenSpec apply — the main agent groups pending tasks by scope and dispatches a worker per group (parallel then serial by dependency), only managing OpenSpec state and reviews; it never writes implementation code itself. Ships /opsx:apply-subagent (Claude subagent worker) with its SOT skill. Requires the agency-agents catalog at ~/.agency-agents (install per the README); a role it cannot resolve there halts the run rather than standing in a fabricated expert.

Portable adversarial review loop for Claude Code: Code Reviewer + Reality Checker are always required; the initial round may add zero or multiple highly relevant Agency Agents domain experts, with no seat cap. It merges findings, applies the smallest accepted fixes, checks and re-reviews them, and gives a repeated blocker one fresh root-cause escalation before NOT-CONVERGED. Missing catalog or structural isolation never blocks the loop; Claude may optionally add a review-only Codex rescue pass. Architecture selection belongs to council.