Use when you need structured retrieval for news, company news, events, announcements, research reports, meeting minutes, morning briefings, timelines, or document details from the Doc Search source namespace.
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
Use when you need structured retrieval for news, company news, events, announcements, research reports, meeting minutes, morning briefings, timelines, or document details from the Doc Search source namespace.
Doc Search
Use this skill for read-only document retrieval from the Doc Search source namespace.
Do not use this skill for structured market prices, financial time series, trading-strategy backtests, direct recommendations, or future event schedules. Descriptive event-level backtest tools in this skill are allowed only for retrieved historical event evidence.
Runtime Binding
This skill defines the Doc Search logical source contract; it does not declare a separate user-visible MCP connection.
In WorkBuddy, call Doc Search tools through the single tongzhou-fin-research Connector. Runtime tool names use doc_search__<tool>, for example doc_search__search_announcements.
In Codex and retained legacy clients, the same canonical tool may appear as a bare name or a client-qualified server/tool name. Match the canonical tool suffix and keep the Doc Search source label.
Do not bypass an available tongzhou-fin-research Connector by selecting similarly named global search tools such as search_hot_news or search_company_news.
Progressive References
Load only the reference needed for the current question:
references/news.md: general news, company news, hot news, morning briefings
Read this skill's references/<name>.md file before filling detailed tool parameters.
For combination queries, read multiple references. Example: "今天新能源为什么大跌" may need references/news.md, references/announcements-events.md, and layer1-fin-data/references/screening.md.
Routing
Choose the narrowest retrieval family first:
The table uses canonical tool suffixes for readability. Under the WorkBuddy Connector, prepend doc_search__ when selecting the actual tool.
Every search call must include at least one non-empty narrowing signal such as query, ticker, company, industry, content_type, source_name, time_window, or date range.
Resolve relative time before calling tools: 今天/昨日/本周/本月/最近N天 must map to a supported time_window, days, or explicit start_date/end_date; include the actual date range in the answer.
For "今天有什么新闻/热点", use search_hot_news(query="今日热点", time_window="1d", limit=10) instead of an empty query.
For company research reports, pass the entity through company and/or ticker in search_research_reports; do not put a company name only in query when using a long window such as 1y.
Each research call accepts one company/listing. Split explicit company/ticker lists into separately labeled calls; MULTI_ENTITY_RESEARCH_UNSUPPORTED is a parameter error, not an empty result.
For a listing-specific research question, preserve the resolved ticker, market, and scope="listing"; do not silently remove ticker or replace listing results with issuer-title matches.
If issuer-wide coverage is useful, make a separate scope="issuer" call and label it as a broader issuer result.
Normal research calls use limit=5-10; the hard single-call maximum is 20.
A listing search may return a short-lived continuation_ref. When more evidence is needed, make the next call with only that reference; the gateway restores the original filters. Continue only while continuation_status="available", and stop once enough evidence is collected or the status is , , or . Treat / as bounded retrieval rather than exhaustive coverage. Older responses may omit ; if they also omit the reference, stop without claiming completeness. Never display, decode, construct, share, or persist the reference.
Output Shape
Prefer compact lists or tables. Include:
title
entity/company/industry when applicable
publish/release date
source or broker
short evidence summary
limitation or data-source caveat when relevant
content_type
source_set
references/limitations.md
list_categories
复杂组合、不确定问法、跨 MCP 解释
references/examples.md
combine the relevant tools
空结果、覆盖范围、实时性、内部字段边界
references/limitations.md
no retrieval tool by itself
complete
limit_reached
unavailable
limit_reached
unavailable
continuation_status
The configured announcement source is CNINFO-oriented and does not include native HKEX disclosures. A Hong Kong announcement empty result is a source-coverage gap, not evidence that the company made no disclosure.
Return title, date/time, source, and short summary first. Do not fetch long body text unless the user asks for details.
Treat snippet as search-match context, summary as a stored source summary, and content as bounded detail evidence. Never relabel one as another.
In get_document, read content_kind before using content: research_viewpoint, body_excerpt, event_analysis, and transcript_excerpt have different evidence meanings. Type-specific fields are under metadata.
summary_status="unavailable" means no stored summary was returned; error means the summary source failed. Do not fill either state from content or model memory.
Use get_document only after a search result provides doc_id and a clear doc_type. For research results, treat doc_id as a short-lived opaque evidence reference: pass it back unchanged, never display or persist it, and rerun the search if it expires.
Use get_document_summaries only for 2-5 explicitly selected search results needed by one evidence task. Pass each returned evidence_ref with its exact doc_type; never construct, persist, mix across users, or use batch summaries as an export path. This tool returns only stored summaries and must not return or reconstruct original text.
Use get_document_source_coverage only when the user asks about source coverage, data time range, or whether an empty result may be a coverage gap. Treat its first/last dates as an aggregate snapshot, not proof of continuous daily coverage.
Search results are candidates for selection. For research, call get_document only for 1-3 selected reports needed by the user's request or claim verification; never expand every result by default.
Do not expose internal collection names, retrieval scores, request plans, or backend routing details to users.
If no result is returned, say no matching documents were retrieved under the current filters. Do not invent events or reports.
Event backtest tools return descriptive historical samples only. Do not present them as a future prediction, strategy backtest, target price, or recommendation.
If a multi-part user question asks for document evidence plus another data domain, finish the document summary and then continue with the other domain before finalizing.
Do not answer "will query / please wait" after tool use. Continue calling the needed tool or provide the completed answer.
Treat a non-empty documents or events list as retrieved evidence. Summarize those items; do not call it an empty result.
For coverage, freshness, timeline, and content-type caveats, read references/limitations.md.
