Web Research Brief
Use this when someone wants researched answers, not a single-link reply: technology comparisons,
claim verification, market/landscape scans, "最新进展" roundups.
You (the agent) run the research yourself. Every claim in the brief must carry a source you
actually fetched. If sources disagree, show the disagreement — do not average it away. Never pad
with prior knowledge stated as if sourced; mark unsourced background explicitly as 背景常识.
How to run it
Scope. Restate the question in one line + 2-4 sub-questions. Time-sensitive topics get an
explicit recency requirement (prefer sources from the last 12 months; date every source).
Resolve a search capability, discovery-first. Use web_search if visible. Otherwise list
connected services with nyxid_services and pick a search/enrichment service by its
endpoint/label semantics (do NOT hardcode slugs), calling it via nyxid_proxy. If no search
service exists, say so, suggest nyxid-service-doctor, and continue with direct web_fetch on
URLs you can justify (official docs, known primary sources).
Search wide, fetch deep. 2-4 queries (different phrasings/angles), then web_fetch the 3-6
most load-bearing results — prefer primary sources (official docs, papers, release notes,
filings) over aggregators. For each: capture title, publisher, date, and the 1-3 facts you take
from it.
Cross-check. Any number, ranking, or strong claim needs 2 independent sources or an explicit
[单一来源] mark. Note conflicts as 来源分歧: A说X(date), B说Y(date) and prefer the newer
primary source.
Compose the brief:
- 一句话结论
- 关键发现 — 3-7 bullets, each ending with
[n] source refs
- 分歧与不确定性 — what is contested or thinly sourced
- 来源 — numbered list: title — publisher, date, URL
Keep it under ~40 lines in chat; if it must run longer, create a Lark doc with
lark_docx_create, send the link, and keep a 5-line summary in chat.
Offer continuity. For recurring topics offer a scheduled_agent_creator watch (cron + this
skill + the standing question).
Guardrails
- Cite or mark: no naked claims. Quote numbers exactly as the source states them, with units.
- Respect recency: never present an undated or old page as current state; always print dates.
- Read-only on the web; deliver only into the current chat (plus the doc link if created).
- Never ask the user for tokens — your NyxID-brokered tools handle all credentials.
1---2name: web-research-brief3version: "1.0"4description: Structured web research from chat — multi-source search, fetch and cross-check, then a compact cited brief (背景 / 关键发现 / 分歧与不确定性 / 来源列表). Uses discovery-first search service resolution plus web_fetch; long reports land as a Lark doc. Use for "帮我调研X / 对比A和B / 这个说法靠谱吗 / 最近关于Y有什么新进展".5metadata:6 source: https://skillmd.com/skills/chronoaiproject/web-research-brief7 category: plain8 tag:9 - aevatar10 - s-capability11 - research12 - web13 - citations14 - brief15---1617# Web Research Brief1819Use this when someone wants researched answers, not a single-link reply: technology comparisons,20claim verification, market/landscape scans, "最新进展" roundups.2122**You (the agent) run the research yourself.** Every claim in the brief must carry a source you23actually fetched. If sources disagree, show the disagreement — do not average it away. Never pad24with prior knowledge stated as if sourced; mark unsourced background explicitly as 背景常识.2526## How to run it27281. **Scope.** Restate the question in one line + 2-4 sub-questions. Time-sensitive topics get an29 explicit recency requirement (prefer sources from the last 12 months; date every source).30312. **Resolve a search capability, discovery-first.** Use `web_search` if visible. Otherwise list32 connected services with `nyxid_services` and pick a search/enrichment service by its33 endpoint/label semantics (do NOT hardcode slugs), calling it via `nyxid_proxy`. If no search34 service exists, say so, suggest `nyxid-service-doctor`, and continue with direct `web_fetch` on35 URLs you can justify (official docs, known primary sources).36373. **Search wide, fetch deep.** 2-4 queries (different phrasings/angles), then `web_fetch` the 3-638 most load-bearing results — prefer primary sources (official docs, papers, release notes,39 filings) over aggregators. For each: capture title, publisher, date, and the 1-3 facts you take40 from it.41424. **Cross-check.** Any number, ranking, or strong claim needs 2 independent sources or an explicit43 `[单一来源]` mark. Note conflicts as `来源分歧: A说X(date), B说Y(date)` and prefer the newer44 primary source.45465. **Compose the brief:**47 - **一句话结论**48 - **关键发现** — 3-7 bullets, each ending with `[n]` source refs49 - **分歧与不确定性** — what is contested or thinly sourced50 - **来源** — numbered list: title — publisher, date, URL51 Keep it under ~40 lines in chat; if it must run longer, create a Lark doc with52 `lark_docx_create`, send the link, and keep a 5-line summary in chat.53546. **Offer continuity.** For recurring topics offer a `scheduled_agent_creator` watch (cron + this55 skill + the standing question).5657## Guardrails5859- Cite or mark: no naked claims. Quote numbers exactly as the source states them, with units.60- Respect recency: never present an undated or old page as current state; always print dates.61- Read-only on the web; deliver only into the current chat (plus the doc link if created).62- Never ask the user for tokens — your NyxID-brokered tools handle all credentials.