deep-research

The saved excerpt is a snapshot from when this skill was listed. The external source remains the complete and most current version.

---
name: deep-research
title: "深度调研"
description: Courses whose content rests on current, external or real-world facts that must be verified against live sources before being taught — recent events, market or policy data, product and version specifics, scientific developments, named cases. Researches the topic first, keeps a claim-to-source ledger, and grounds the outline and every page in what was actually fetched. Use when the request depends on up-to-date or externally checkable facts; not for timeless textbook topics that stand on established knowledge alone.
---

# Deep Research course design

You are designing a course whose content rests on facts you must **verify,
not recall**. Research first, outline second, generate third. A number, date,
name or finding enters the course only because you saw it in a source you
fetched or in the user's own material — and you can say which one.

## Structure

- Open with one `slide`: it frames the research question and previews what
  kind of evidence the course will examine. Not a definitions or
  history-of-the-field page.
- The body carries the findings, in whatever scene types fit: slides for
  sourced exposition, `interactive` for evidence the learner can inspect,
  `quiz` for checking whether the learner can tell a supported claim from an
  unsupported one.
- Close on what the evidence establishes and where it runs out — not on a
  generic summary.

## Step 1 — Start from what the session already has

Call `list_materials` before any search. Materials the user attached —
documents, links, data, recordings — are the primary authority on their own
subject; web research supplements them, it does not replace them. If a
derivative is still extracting, `extract_material` or `wait_for_materials` as
in any other course. URLs the user pasted in chat can be fetched directly with
`fetch_url`.

## Step 2 — Split the topic into facets

Break the request into 2–4 searchable facets — distinct questions the course
must answer with evidence. Typical facets: current state or latest
developments; authoritative figures and baseline data; concrete cases and
incidents; risks, controversies or open questions. Write the facet list down
before searching. Not every facet needs a search: a facet that is stable
textbook knowledge is skipped and taught as such.

## Step 3 — Search with a budget

- At most **8 `web_search` calls for the whole session**, planned across the
  facets. One precise query beats several vague ones; write queries in the
  language of the course.
- The session shares one run with planning, `set_roster` and
  every page's generation, and every extra call is latency the user watches.
  Research is one slice of the run, not the main act. If the run gets long,
  cut facets — never generation.
- Read each result before searching again: the next query should be shaped by
  what the last one returned, not a rewording of it.

## Step 4 — Pick sources, fetch them

- From the search results, pick at most **6 URLs in total** across all facets.
  Prefer primary and authoritative origins — official bodies, named
  institutions, the report or dataset itself — and for time-sensitive claims
  prefer the most recent. Skip mirrors and aggregators repackaging the same
  story: fetch one origin, not three copies of it.
- `fetch_url` accepts only URLs that appeared in the user's messages or in
  this session's `web_search` results. Never assemble or recall a URL from
  memory — if the source you want did not surface, refine the search instead
  of guessing an address. A URL that never surfaced does not exist for this
  course.
- `fetch_url` ingests the page as a session material and returns a
  `materialId` plus a first-page preview. That `materialId` is what the ledger
  cites.

## Step 5 — Read deep, keep the ledger

- Page through each fetched material with `read_material` — at least far
  enough to verify every claim you plan to take from it. Use `search_material`
  to locate a specific figure or na