research

name

research

description

Dispatch the Research Team on any question. Amplifies a terse prompt into a full multi-specialist investigation with adversarial gates (skeptic + adversary + moderator + evaluator). Use this skill IMMEDIATELY when the user asks to research something, investigate a topic, check what's happening with a technology, compare options, or asks "what's the SOTA on X". Examples: "research vLLM MoE routing", "check HN about agent memory", "what's the best approach for X", "investigate why Y is slow", "compare A vs B vs C".

Research Team dispatch

You are invoking the Research Team (v2.1 protocol, 17 specialists).

What to do

1. Read ~/.claude/agent-memory/research-lead/MEMORY.md (first 200 lines) — these lessons are binding on dispatch decisions.

2. Read ~/.claude/agents/research/research-lead.md — adopt this as your behavioral contract for this session.

3. Create the session workspace at <cwd>/.claude/teams/research/<slug>/ where <slug> is derived from the user's question (e.g. "vllm-moe-routing-2026q2").

4. Execute the full v2.1 protocol:

   • Round 0: QUESTION.md + HYPOTHESES.md + EXPECTED_EVIDENCE.md + planner dispatch
   • Round 1: Wide opener (6-10 specialists in parallel)
   • Round 2: Synthesist + moderator (on contradictions) + skeptic + adversary
   • Round 3: SYNTHESIS.md draft + evaluator 5-dim rubric
   • Close: Retrospector + scribe

5. Run audit gates:

   • Before synthesist: python3 ~/.claude/scripts/audit_evidence.py <slug> --gate=mid-flight
   • Before SYNTHESIS: python3 ~/.claude/scripts/audit_evidence.py <slug> --gate=synthesis --strict

6. Deliver the synthesis to the user.

Hard rules

• All Opus, all max effort
• Every specialist must produce an evidence file (full-activation enforcement)
• Adversary is mandatory if any web/community sources are cited
• The 14-day freshness sweep is mandatory for fast-moving topics
• MEMORY.md lessons are binding, not advisory