Standardmäßig ist der Prompt ausgewählt, der zuerst die Quelle prüft. Sie können zu einem direkten Befehl wechseln oder eine lokale Kopie herunterladen.
Quelldateien prüfen
Lesen Sie SKILL.md und alle von SkillsMP angezeigten Begleitdateien, bevor Sie sich für eine Installation entscheiden.
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Ein direkter Befehl überspringt den Prüf-Prompt. Prüfen Sie die Quelle, bevor Sie ihn ausführen.
Preferred for a named security's earnings preview; includes quote, profile, earnings, estimates, targets, ratings, financials, coverage, and current-listing evidence
data_fetch(data_type="quote", ticker="ticker")
Fetch detailed latest-available quote data such as price, change, volume, and provider timestamp
data_fetch(data_type="profile", ticker="ticker")
Fetch company details such as business description, industry, and CEO
Fetch ETF/fund holdings after profile confirms isEtf or isFund
data_fetch(data_type="news", ticker="ticker")
Fetch current-turn news for the exact security
data_fetch(data_type="gainers_losers")
Broader market scan when a screening request needs candidates
data_fetch(data_type="sector_performance")
Sector strength context for screening or relative positioning
web_search(query="...")
Search for news, analyst views, and recent events
Adapt To The Requested Outcome
Read the complete request and choose the evidence and answer shape that best fits it; these are reusable answer patterns, not a closed intent classifier or a grammar that the user's wording must match:
Research mode: the user asks about one company, ETF/fund, ticker, fundamentals, technicals, or recent developments
Valuation mode: the user asks whether a company looks rich, cheap, stretched, fairly priced, or wants a valuation bridge / peer view
Screening mode: the user asks for a shortlist that matches factors such as AI, dividend yield, value, growth, or momentum
Non-negotiable Current-turn Pipeline
In the main agent loop, read the complete current user query and retain every possible named security before answering. Treat any pre-scanned ticker as a candidate seed, never as proof that the entity set is complete. Do not rely on a service-side grammar that splits natural language on commas, and, 和, &, or /; names and symbols such as AT&T, S&P Global, BRK/B, and Berkshire Hathaway, Class B must remain understandable in context. Assign each named security one stable, distinct, case-sensitive entity_route key for this run and reuse it verbatim. Start with one batch/parallel discovery round using one separate data_fetch(data_type="search", query="...", entity_route="...", identity_match="...") call per named security; this tool-enabled round returns only tool calls, never a data-time line, summary, draft, or final prose. Calls may run in parallel, but never combine multiple securities into one query. identity_match is call-scoped and must be present on every search: set identity_match="exact_symbol" when the query is a ticker and identity_match="name_or_alias" when it is a company name, Chinese name, or alias. A previous declaration does not authorize a later search, and the service must not guess from case, length, or punctuation. Reuse the exact same route key on that security's refinement, quote, profile/snapshot, and later DataFetch calls. An exact-symbol route keeps its symbol constraint even during a later company-name refinement, with only bounded provider separator equivalence such as BRK/B, BRK-B, and BRK.B; an ETF or product whose name merely embeds CRWV cannot replace CRWV. The key is internal evidence linkage, not a user-visible entity claim. If a Chinese name or alias search is empty, refine it within the same route and use refines_query to copy the original empty query verbatim with matching case. If an earlier search omitted the route key, repeat its exact query or set supersedes_query to that old query verbatim with matching case so at most that one provisional route is migrated; never guess alias equivalence or erase another entity's gap. refines_query and supersedes_query are strictly mutually exclusive: provide at most one on any search, because providing both invalidates that identity search and leaves the route pending. After search results return, select one standard symbol per route and give that same symbol both quote and profile/asset-route coverage. A plain ticker such as NBIS, , , or is normal user input: query it directly instead of asking for a company name. Only ask for clarification after current-turn tools still show genuine ambiguity or no authoritative coverage.
Research Mode
Resolve every named security discovered from the complete current query with current-turn tools, preferably in one batch/parallel first round. A ticker is a first-class search input but becomes an entity only after exact-symbol confirmation; names, aliases, Chinese names, multiple securities, and share classes must all produce explicit resolution results. A pre-scan miss must fall through to this agent loop, not become a user-facing failure. Never take the first approximate result silently, and clarify only when tool evidence remains genuinely ambiguous.
Verify the current-turn same-symbol quote, then select the route from structured exact-symbol evidence. A company uses profile, financials, and news; an ETF/fund confirmed by profile isEtf/isFund uses etf_holdings and news; a crypto asset confirmed by exact search market evidence such as exchangeShortName=CRYPTO uses the same-symbol quote and relevant news. Never request corporate financials or an earnings calendar for a confirmed ETF/fund, and never request corporate financials, an earnings calendar, or ETF holdings for crypto. Treat provider errors separately from a successful empty response. Do not infer an asset type from an empty response.
A quote-only question may stay concise. A deep single-company, quarter-outlook, “can it take off”, fundamentals, valuation, or buyability question must use these nine numbered sections in order:
Conclusion
What the company is and how it makes money
Moat and competitive barriers
Industry position and key competitors
Financial quality
Valuation using at least two suitable methods only when the current-turn inputs are complete; otherwise use the method that can be calculated rigorously and state the missing inputs
Preserve the Agent-authored first-line data timestamp and quote basis, and distinguish verified facts, inference, conclusion, and action. Do not ask for the user's cost basis as a substitute for completing the analysis.
If required live evidence is missing or mismatched, stop numeric conclusions instead of filling gaps from memory, history, profiles, or another symbol.
If the user explicitly asks for a chart, trend line, comparison visual, or the answer would be materially clearer as a chart, hand off to chart_visualization with the concrete numbers you already fetched.
If the exact quote and profile are valid but current company financial statements are empty, failed, mismatched, or limited to an income statement, do not fail the whole response and do not fabricate values. Keep all nine sections, state 本轮公司财务数据未核验 in section 5 with the exact missing scope, and base the remaining sections only on verified quote/profile/news evidence. An income statement does not prove cash, debt, net debt, or free cash flow. Financial-data absence must never be rewritten as an absence of current quote capability.
ETF / Fund Research Route
When the exact-symbol profile confirms isEtf=true or isFund=true, use these nine numbered sections instead of the company template:
Conclusion
Fund objective, strategy, and tracked exposure
Holdings, concentration, and primary exposures
Geographic, sector, and currency risk
Liquidity, fund size, and trading characteristics
Fees, tracking error, and underlying-asset valuation framing
Preserve the Agent-authored first-line data timestamp and quote basis, and separate verified facts from inference and action. If holdings, fees, size, or tracking-error evidence is absent, label that item as not verified in the current turn; do not fill it from memory. A successful empty company financial response for a confirmed ETF/fund is not a provider outage and must not block this route.
Crypto Research Route
Only classify crypto from exact-symbol structured market evidence such as exchangeShortName=CRYPTO; do not infer it from a USD suffix. A confirmed crypto asset uses quote and relevant news, not stock profile, company financials, an earnings calendar, or ETF holdings. Use nine substantive numbered sections: conclusion with verified current price; asset/network/use case; supply/tokenomics/concentration; adoption/liquidity/market structure; on-chain/network/ecosystem evidence; valuation framework and assumptions; Bull/Bear/Base; catalysts/regulation/risks/falsification; and an action with trigger conditions. Label absent on-chain, supply, or ecosystem evidence as not verified in the current turn.
Valuation Mode
Resolve the ticker first, fetch the same-symbol quote, and read the exact-symbol profile; do not attempt valuation before confirming whether the entity is a company or an ETF/fund.
For a company, fetch financials; add quote or snapshot if you also need current market context. Use at least two suitable methods only when every numerator, denominator, period, and balance-sheet input is present. Annual FY revenue is not TTM. Without verified cash/debt or enterprise value, label market-cap/EBITDA as such and never call it EV/EBITDA. If only one method is fully supported, use that method, disclose the missing inputs, and do not invent net debt, historical multiples, target prices, or technical support levels to fill the template.
For an ETF/fund confirmed by isEtf/isFund, fetch etf_holdings plus quote and frame valuation through underlying holdings/exposures, fees, tracking error, concentration, and applicable portfolio-level multiples. Do not fetch corporate financials or an earnings calendar, and do not apply a single-company DCF to the fund itself.
Use web_search for the latest operating updates, strategy changes, holdings disclosures, guidance changes, or peer-comparison context appropriate to the confirmed asset type.
Do not collapse the result into a simplistic categorical verdict with no assumptions attached.
Screening Mode
Extract the user's explicit criteria before naming companies
Use gainers_losers, sector_performance, or targeted web_search to form an initial candidate set
Narrow the result to 3-5 names, exact-resolve every final candidate, and fetch a same-symbol snapshot or quote for each
Return a comparison shortlist with why each name matches the screen, plus the main risk or diligence gap for each one
Do not output a blunt recommendation list without comparison logic or caveats
INTL
RMBS
CRWV
After identity is confirmed, fetch the same-symbol quote and preserve its provider timestamp. Never establish identity, price, change, financials, or news from assistant history or model memory.
Select the company, ETF/fund, or crypto route only from current-turn structured evidence. A named security takes precedence over broad market words in the same query.
When hone_security_listing_evidence.status=active_listing, treat the same-symbol security as currently listed and trading. Never override it with model memory about an earlier acquisition, delisting, or former parent, and never redirect the user to that former parent. If current regulatory evidence conflicts with the provider, fetch and disclose the authoritative conflict instead of deciding from memory.
Interactive final-answer ownership stays with the main Agent and the ordinary function-calling loop. Current-turn business-tool results remain directly in the same Agent context. Before the entity/evidence floor is complete, return only the missing real tool calls. After that floor, either call another real business tool when the user's actual question still lacks key evidence, or return one complete natural Stop + Done final answer. Tool-enabled rounds remain buffered because a provider can emit prose before a later tool call; discard that preamble when tools appear, and publish/persist only the one completed DirectFinal. The Agent itself must emit 数据时间：北京时间 YYYY-MM-DD HH:MM；行情口径：... as the first visible line, using the current Beijing time from the Session context. Quote time, price, and market basis must come from current-turn quote fields. Prefer hone_quote_time.beijing; market_date_new_york and new_york only represent New York local date/time and never prove an exchange, a market session, or a closing price. Take the exchange only from exchange / exchangeShortName, include a session only when a tool explicitly verified it, and do not emit a preamble before the first line.
Use absolute-date web_search for current events, causes, policy, analyst context, customer/supplier relationships, contracts, purchase scale, ownership, or competitive claims. data_fetch(search) proves only the entity candidate and profile proves only the company's business description; neither proves a relationship or news causality. For a broad “A and B relationship” question, let the Agent derive relevant axes from the complete semantics; normally investigate commercial/customer-supplier/technology-contract and investment/ownership separately, preferably in parallel and through SEC, company IR, or both parties' announcements. One generic query is not complete research. A search snippet may support only the limited fact it explicitly states; never expand it into an unstated contract change or cause, and disclose when full text or a primary source was not verified.
Before finalizing, reread the exact current-turn evidence snippets/fields. Every relationship fact's number, direction, rank, role, right/obligation, product model, and valuation label must occur directly in current-turn evidence. A URL locates a source but does not prove unsupported text. Render any judgment beyond literal sources as a separate sentence beginning Inference:; when the premises are insufficient, delete it. In particular, do not summarize a relationship as core/largest, a major customer, highly dependent, locked in, or multiply bound unless the current source text directly supports that strength and direction. Relationship answers should be minimal and on-scope rather than filling an unrelated deep-company template.
When a same-symbol quote succeeded, never claim that real-time/current market data was not requested, unavailable, or outside Hone's capability. Describe it accurately as the latest available provider quote, not tick-by-tick data.
