A Share Site Crawl

Crawl and validate A-share data from five key sites using browser or fetch methods, handling access limits and normalizing market-news for reliable summaries.

Security Scan

The skill's instructions, references, and requirements are coherent with its stated purpose (crawling and normalizing A-share content); it is instruction-only, requests no credentials by default, and contains sensible downgrade/verification rules — but exercise caution before supplying any login cookies or a logged-in browser relay.

Purpose & Capability

Name/description match the internal content: all files and SKILL.md focus on crawling and normalization for the five A-share sites. There are no unrelated env vars, binaries, or install steps requested, and the included reference docs align with the stated objective.

Instruction Scope

Runtime instructions are detailed and scoped to crawling, classification, normalization, and downgrade behavior. They explicitly instruct reading the local reference files bundled with the skill. They also note optional requests for stronger access (Chrome relay tab, logged-in profile, cookies) when the user explicitly asks — this is conditional scope expansion the user should be aware of.

Install Mechanism

No install spec and no code files — instruction-only skill. This minimizes disk write/execute risk; nothing is downloaded or installed by the skill itself.

Credentials

The skill declares no required env vars or credentials which is proportional. However, the instructions advise asking for authenticated artifacts (cookies, logged-in browser relay) to improve access for some sites; providing those would be highly sensitive and should be explicitly consented to by the user.

Persistence & Privilege

Flags are default (always:false, user-invocable:true, autonomous invocation allowed). The skill does not request permanent presence or modify other skill/system configs. Autonomous invocation is standard and not flagged by itself.

Assessment

This skill is internally consistent and appears to do what it says: probing five Chinese A-share sites, classifying access quality, and normalizing records. It is instruction-only and asks for no credentials by default. Key cautions before you install or use it: - Do not provide login cookies, browser profiles, Chrome relay tabs, or other credentials unless you explicitly want deeper, authenticated extraction — these are sensitive and can expose your account. - The skill may ask you for stronger access in order to bypass anti-bot or login walls (e.g., for 雪球). Only grant that access deliberately and understand the privacy/security implications. - The skill references no external upload endpoints or installers, but the source is listed as unknown; if you plan to run this in production, confirm provenance and terms-of-use for the target sites (some scraping may violate site terms or rate limits). - Because it can be invoked autonomously by the agent (normal default), consider whether you want to allow automatic periodic crawls — if not, restrict invocation or require explicit user confirmation for cron-style operations. If you want stronger assurance, ask the publisher for: (1) the skill's author identity and license, (2) any telemetry or outbound endpoints the skill would use if extended, and (3) an explicit statement about how it would request/handle any credentials you might supply.

Use this skill to collect public A-share information from the five target sites and to convert raw site access into repeatable summary-ready records.

Read Order

Always read these first:

• references/sites.md
• references/workflow.md

Read these in addition when the task involves formal collection, normalization, or recurring jobs:

• references/entrypoints.md
• references/fields.md
• references/risks.md

Use references/entrypoints.md for fixed site entry pages, verification status, cron priorities, and default crawl mode.

Use references/fields.md for the normalized schema, source tiering, credibility, opinion-risk handling, content typing, cron retention, time normalization, ticker normalization, and dedup rules.

Use references/risks.md for P0/P1/P2 risks, recognition signals, and downgrade or mitigation decisions.

Core Rule

Prefer browser for page truth and web_fetch for cheap probing.

• Use web_fetch first when the site is known to have stable public text pages
• Use browser first when the site is dynamic, disclosure-driven, or clearly stronger in rendered form
• If both fail, report the site as restricted or missing instead of pretending it was covered
• Do not treat anti-bot code, disclaimers, shells, or login walls as usable content

Working Workflow

1. Start from the correct page type

• Prefer fixed entrypoints, list pages, search pages, disclosure pages, telegraph streams, and stock-detail pages
• Do not judge 巨潮资讯 from homepage-only text
• Do not rely on noisy portal homepages when a better inner page exists

2. Probe and classify access

Judge each probe into one of these buckets:

• usable: readable and materially sufficient
• partial: some content is real, but clearly incomplete
• shell-only: mainly navigation, scripts, disclaimers, or boilerplate
• blocked: anti-bot, login wall, or meaningless payload

3. Choose extraction mode

Use one of these verdicts per site or page:

• fetch-first
• browser-first
• restricted
• not-usable

4. Keep site roles distinct

• 巨潮资讯: official confirmation and disclosure verification
• 东方财富: public aggregation, data-center navigation, and quasi-structured market pages
• 财联社: fast market events and telegraph flow
• 韭研公社: topic logic, timeline, and community clue discovery
• 雪球: sentiment, heat, stock-detail snapshots, and community discussion

5. Normalize before summarizing

When the task is more than a one-off crawl check, convert findings into normalized records using references/fields.md.

Minimum normalization discipline:

• assign source_tier, credibility, content_type, and opinion_risk
• normalize time to Asia/Shanghai when possible
• normalize A-share tickers conservatively
• deduplicate repeated event coverage
• separate confirmed facts from market claims and sentiment

6. Apply downgrade rules early

Use references/risks.md when deciding whether to downgrade, defer, or replace a source.

Default downgrade behavior:

• login-gated or anti-bot content -> restricted
• shell-only or disclaimer-heavy result -> switch entrypoint or switch tool
• 财联社 telegraph 默认先保留列表正文; only hit detail when the list is truncated, a canonical URL is needed, or an original-source jump matters
• 巨潮公告默认先保留列表元数据; only chase PDF when the title is high-value enough to justify body extraction, otherwise keep title-derived summary and mark that PDF body was not extracted
• community-only claim without confirmation -> keep as clue, not fact
• unavailable priority site -> disclose it and use approved fallback public sources

Default Site Priority

Use this order for stable public collection when the task does not specify a scenario:

1. 东方财富
2. 财联社
3. 巨潮资讯
4. 韭研公社
5. 雪球

This order reflects public accessibility and extraction stability, not market importance.

When to Ask for Stronger Access

Ask for stronger access only when the user explicitly wants better extraction from a restricted site, especially 雪球.

Examples:

• attached Chrome relay tab
• logged-in browser profile
• cookies or authenticated environment
• a dedicated crawler or site-specific script

Scenario Call Contract

When a cron or caller specifies one of these scenario ids, treat it as a compact instruction bundle and do not ask for a longer prompt:

• pre-open: read references/entrypoints.md, references/fields.md, and references/risks.md; use the pre-open priority order; focus on overnight macro or overseas linkage, policy or industry catalysts, key announcements, expected hot sectors, and today's watchlist
• midday: read references/entrypoints.md, references/fields.md, and references/risks.md; use the intraday priority order; focus on morning index and turnover snapshot, leading or lagging themes, style or sentiment shifts, active stocks with catalysts, and deviation from the pre-open setup
• late-session: read references/entrypoints.md, references/fields.md, and references/risks.md; use the intraday priority order; focus on whether the afternoon main line strengthens or rotates, late-session anomalies, money-flow return direction, hot-stock persistence, and signals that may affect post-close review or next-day expectations
• post-close: read references/entrypoints.md, references/fields.md, and references/risks.md; use the post-close priority order; focus on index and turnover recap, main-line review, key stocks and drivers, important announcements plus exchange or regulator dynamics, and next-day clues with risks

For every scenario:

• keep the output in Chinese and lead with conclusions before detail
• keep 已确认事实, 市场观点与情绪, and 待核实线索 clearly separated
• keep 本轮缺失站点 and 来源层级说明 in the final output
• bind every round to the entrypoint, field-normalization, and risk-downgrade rules instead of freehand summarizing
• do not output buy or sell recommendations

Standard Output

When producing a formal round output, always structure it with at least these sections:

• 已确认事实
• 市场观点与情绪
• 待核实线索
• 本轮缺失站点
• 来源层级说明

Use the sections as follows:

• 已确认事实: only T1 or well-supported T2 items, or items clearly marked as partially confirmed
• 市场观点与情绪: T3 discussion, heat, consensus drift, and sentiment signals
• 待核实线索: rumors, single-source community claims, partial clues, or conflicting statements
• 本轮缺失站点: blocked, unstable, login-gated, or otherwise uncovered priority sites and what fallback was used
• 来源层级说明: explain T1/T2/T3 usage and remind the reader that community sources are not equal to formal disclosure

Per-Site Quick Output for Crawlability Tasks

When the task is specifically about site feasibility rather than a market summary, return:

• Site
• Status
• Recommended mode
• Best entry page
• What works
• Main limitation
• Next step

Non-Negotiables

• Distinguish confirmed facts from community opinion
• Prefer official disclosure and high-confidence public reporting over discussion boards
• Do not output buy/sell recommendations
• Do not imply full coverage when a priority site failed or was inaccessible