deep-research

这段内容是审核时保存的快照。外部来源才是完整且最新的版本。

---
name: "deep-research"
description: "Run a disciplined, multi-source research investigation for a high-stakes question or decision — fan-out web search across many channels, parallel sub-agents, source triangulation (each claim backed by ≥3 independent sources), an adversarial review pass, and every source saved to its own file with verbatim quotes for reuse. Use when a low-quality answer is expensive: strategy work, comparing N products/methods/markets, validating a hypothesis with external data, or mapping how a field works. NOT for quick fact-checks (answer directly), structured 12-dimension competitor scoring (use competitive-teardown), or fast topic overviews where the decision risk is low (use the research router instead)."
---

# Deep Research — Disciplined Meta-Research

Turn "research this topic" into an auditable, reusable investigation instead of a one-shot wall of text. The output is a folder you can return to in a month: every claim traces to a specific source file, the plan documents *why* each choice was made, and a refresh protocol lets you update it later without re-running everything.

**This is the heavy, methodical end of research.** It is not a fast overview — it is the workflow you reach for when getting the answer *wrong* costs more than the tokens spent getting it right.

## How it differs from a quick research router

A router-style research skill (keyword-classify → delegate → short sequential search → markdown brief) is optimal when you need an answer fast and the decision risk is low. `deep-research` is the opposite trade: it pays for rigor. Use it when the answer feeds a strategy, an irreversible decision, a published artifact, or a hypothesis you need to actually test — situations where a shallow fallback would be a liability.

Concretely, `deep-research` adds what a fast overview does not: falsifiable hypotheses up front, parallel sub-agent fan-out across many channels, triangulation with explicit source-type diversity, a mandatory adversarial pass, per-source files with verbatim quotes, and a `refresh_targets.md` for delta-updates later.

## The pipeline (9 phases)

Depth scales with the task — `shallow` runs the core phases inline; `medium`/`deep` add capability discovery, verification, and refresh targets.

| # | Phase | What it does |
|---|-------|--------------|
| 1 | **Reframe** | Rewrite the question, fix the underlying decision, state 2–4 *falsifiable* hypotheses |
| 2 | **Genre & blocks** | Pick the report genre (qa / explainer / decision / landscape / validation / custom) and its building blocks |
| 3 | **Plan** | Write `plan.md`: scope, structure, sourcing strategy, opposition queries, risk register, stop-criteria |
| 3.5 | **Capability discovery** | Audit available API keys/channels in the environment; map subtopics to sources; fall back to HTML where needed |
| 4 | **Search** (loop) | Dispatch sources → launch sub-agents in parallel → fetch & dedup → save each to `sources/NN.md`; re-evaluate between rounds |
| 5 | **Score & triangulate** | Rate every source on Credibility / Recency / Bias; require ≥3 independent, differently-typed sources per thesis |
| 6 | **Synthesize + adversarial** | Assemble the report from blocks, run 4 self-critique questions, add steel-manned counter-arguments |
| 6.5 | **Verify** | Lightweight citation check before closing |
| 7 | **Refresh targets** | Extract entities / numbers / hypotheses into `refresh_targets.md` — the entry point for future updates |

## Core mechanisms

These are what separate a documented investigation from a confident guess:

- **Triangulation.** Every thesis must be backed by ≥3 independent sources of *different types* (primary / academic / industry / discussion). A claim with fewer is flagged "insufficient evidence," not stated as fact.
- **Source-grounding.** Each source becomes its own `sources/NN_slug.md` with metadata, verbatim quotes, and scores. No dangling claim — every assertion links back to a specific file. An empty fe