claude-orchestra

Classify and file every newly installed skill, plugin, MCP, connector, or agent into the right orchestra — never archive. Use whenever the user installs/adds a new repo, skill, plugin, MCP server, or agent ("I installed X", "added Y", "npx skills add Z", "new repo", "set up this MCP", "just got this connector"). Runs the security scan, classifies the tool, places it in an orchestra with usage notes, logs the assignment, and reports back. This is the self-organizing layer of the orchestra system.

Runtime router for the orchestra system (v3 lean doctrine). Invoke when genuinely orchestrating multi-agent work or when unsure which specialist domain fits — build, design, research, marketing, content, SEO, leads, product, video, analytics, knowledge, documents, ads, automation, AI/ML, iOS, planning, growth, investing, cybersecurity — or for ideation/business-planning (which escalates to the NEXUS meta-conductor). Reference ~/.claude/rules/orchestra-system.md for the full roster.

Use when an orchestra has multiple skills, MCPs, or connectors that could plausibly handle the same task (e.g. frontend has frontend-design + ui-ux-pro-max + remotion + Figma + 21st.dev). Scores each candidate against the user's vision, prior preferences, and task complexity, picks the winner, EXPLAINS the choice when non-obvious, and asks the human to disambiguate when confidence is below 70%. Logs every pick to learnings for pattern improvement over time.

Use BEFORE any orchestra activation AND BEFORE any completion claim. Enforces Karpathy's 4 principles, Superpowers' Iron Law (no completion without fresh verification evidence), and ASK_HUMAN as a first-class status. Triggers on every entry to an orchestra, every handoff between orchestras, and every user-facing success claim ("done", "works", "should pass", "great", "perfect"). The orchestra system's anti-hallucination wrapper.