deep-researchlisted

# Deep Research Create high-fidelity research reports with strict format control, evidence mapping, source governance, and multi-pass synthesis. ## What makes this produce provider-grade output Three things drive deep-research quality, in order of impact. This skill is built around them; do not skip them under time pressure. 1. **Compute spent through a loop** — dozens of queries and full-page reads, not 5-6 snippet searches. Under-spending is the #1 quality leak. 2. **Dynamic re-planning** — decide what to search NEXT from what you just READ. The initial plan is a hypothesis, not a contract. See [references/replanning_loop.md](references/replanning_loop.md). This is the difference between this harness and static RAG. 3. **Verification** — a pass that checks every claim maps to a source that actually supports it. See P6/P7. Semantic search / embeddings are NOT a quality driver for open-web research — a search engine already ranks the open web. Reranking belongs only on a closed corpus, or as a last-mile token-budget trim on already-scraped pages. See [references/retrieval_tiers.md](references/retrieval_tiers.md). ## Architecture: Lead Agent + Subagents + Replan Gate ``` Lead Agent (coordinator — minimizes raw search context) | P0: Probe environment (tier) + set effort budget + source policy | P1: Initial decomposition into sub-questions (HYPOTHESES, not a frozen plan) | P2: RESEARCH LOOP ───────────────────────────────────────┐ | Dispatch ──→ Subagent