qveris-anthropic-financial-services

v2026.09.25

QVeris-native adaptation of candidate 1, Anthropic Financial Services. Use for institution-style earnings review, peer comps, DCF assumption audit, model update, and market research workflows that must use qveris_finance.* CAP tools with trace-backed evidence and no investment advice.

GitHub
安装命令
npx skhub add qverisai/qveris-anthropic-financial-services
Markdown
SKILL.md

QVeris Anthropic Financial Services

Use this skill to create institution-style equity research artifacts adapted from Anthropic Financial Services while replacing all source data access with QVeris finance CAP tools. Keep the original business shape: earnings analysis, comps, DCF/model updates, and market research memos.

Source record:

FieldValue
Candidate number1
Original repositoryAnthropic Financial Services
GitHub URLhttps://github.com/anthropics/financial-services
LicenseApache-2.0
Evaluation recent activity2026-06-26
Local source snapshotthird_party/source_repos/01-anthropic-financial-services
Snapshot latest commit4aa51ed on 2026-06-26

Runtime Contract

  • Use only qveris_finance.* CAP tools for financial data.
  • Use only QVERIS_API_KEY; do not request, store, or infer third-party keys.
  • Resolve entities first with qveris_finance.ref_symbology, qveris_finance.ref_security_master, and qveris_finance.ref_company_profile.
  • Accept dry_run, max_calls, max_age, and budget_note; if omitted in a natural-language request, default to dry_run=false, no hard max_calls limit, max_age=P1D, and a conservative budget note, then echo those controls.
  • Use the shared retry policy at ../references/qveris-finance-retry-policy.md; retry transient 5xx/transport failures at most 2 times, do not blind-retry 404s, and hard reject semantic mismatches.
  • Include qveris_trace for every numeric claim, quote, event, and conclusion.
  • Report missing_fields; do not fill missing data as fact.
  • Treat QVeris _meta.source_provider as internal provenance only; never call, request credentials for, depend on, print, or copy it into sanitized trace.
  • Normalize trace provenance: qveris_trace[].tool_name and any human-readable trace labels must use only qveris_finance.* capability names. If QVeris returns vendor/provider IDs in _meta.source_provider or _meta.failover_log, drop those fields from sanitized trace and describe only the QVeris capability status, validation result, fallback, and missing fields.
  • In final user-facing output, do not name external providers even when explaining prohibited fallbacks; say "non-QVeris sources" or "external provider routes" instead.
  • Suppress analyst_target_price, target_price, price-objective, upside, buy/sell, and recommendation fields even if a QVeris payload contains them.
  • Sanity-check entity, market, date window, fiscal period, and payload shape before using data; if a payload is stale, cross-period, truncated, or semantically mismatched, mark it in data_quality and missing_fields.

Evidence Gate

Read ../references/qveris-finance-data-quality-rubric.md before using QVeris payloads as evidence. A payload that succeeds transport but fails identity, date-window, fiscal-period, or statement-consistency checks is hard rejected, not treated as a usable fallback.

  • Use evidence status labels from the shared rubric: complete, partial, proxy_only, or insufficient.
  • Require IS, BS, CF, and segment data to share fiscal year, fiscal period, and period ending before building aligned earnings tables.
  • If a requested annual/FY statement such as FY2025 cash flow returns a latest-quarter or TTM-shaped payload, treat the current call as period_mismatch, not as requested-period evidence.
  • After a period_mismatch, inspect cap-detail and retry once with stricter documented fields such as fiscal_year, fiscal_period, period_type, period, or limit; if the payload still does not match, mark the requested statement missing, for example FY2025 cash flow missing due to period mismatch.
  • Exclude CF from aligned tables when same-period CF net income materially conflicts with IS net income; mark statement_semantic_mismatch.
  • Treat qveris_finance.news_fin_tagged as qualitative context only when transcript or cluster routes are unavailable. Do not derive management quotes, numeric sentiment, strong catalysts, or strong risk direction from tagged news alone.
  • Keep invalid, failed, rejected, unavailable, or weak-relevance CAPs out of Evidence Used. Put them only in Data Quality And Missing Fields, missing_fields, or Trace Appendix with reason codes such as capability_unavailable, semantic_mismatch, period_mismatch, entity_mix, weak_relevance, or insufficient_observations.
  • Use qveris_finance.research_analyst_reports only after issuer and document-type relevance checks pass. Reject academic papers, technical articles, broad industry pages, and weak text matches as weak_relevance; do not treat them as sell-side analyst evidence.
  • Summarize long QVeris payloads in full-workflow reports. If a response is truncated or too large for a compact table, mark payload_summarized or payload_truncated and offer a single-capability note for inspection.

DCF Input Gate

  • Allowed DCF inputs: validated entity/profile, market quote, aligned IS/BS/CF rows that pass fiscal-period and statement-consistency checks, fresh FX, and explicitly returned consensus fields.
  • Conditional DCF inputs: stale or monthly rates only as lagged proxy assumptions, manual trailing ratios only when source fields are present, and tagged news only as qualitative risk context.
  • Denied DCF inputs: period-mismatched CF, CF excluded by statement mismatch, forward multiples inferred without consensus or derived-ratio evidence, analyst target/upside fields, buy/sell language, and growth assumptions derived from tagged news alone.
  • When denied inputs are present in QVeris payloads, list them in Data Quality And Missing Fields and leave the model input blank instead of filling a placeholder.

CAP Invocation

  • Use native qveris_finance.* tools only when that runtime applies the same Skill-owned adapter and returns a qveris.finance-parameter-adaptation.v1 audit; otherwise use this Skill's CLI.
  • If native functions are not exposed but the repo script is available, execute standardized CAP calls from the repository root with node {baseDir}/scripts/qveris_finance_tool.mjs cap-query qveris_finance.<capability_name> --param key=value --safe-json. Use repeatable --param flags for shell-safe parameters; reserve --params '<json>' for complex nested payloads.
  • Treat the Skill-owned CLI as the mandatory finance adapter: it resolves the live canonical CAP, filters and losslessly converts parameters, never copies sample values, permits at most three fully audited attempts, and rejects success=false, missing required fields, wrong entity/market/date/period, and stale real-time data. Use only its qveris.finance-parameter-adaptation.v1 audit and actual attempt parameters in Trace.
  • Direct HTTP CAP calls are transport implementation details of scripts/qveris_finance_client.mjs; workflow code must not bypass the Skill-owned adapter.
  • Use cap-search or GET /capabilities/search only when the CAP ID or parameter contract is uncertain; use cap-detail or GET /capabilities/{capability_id} to verify fields.
  • Generic QVeris discovery, /tools/execute, raw provider routes, and legacy finance shims are prohibited fallbacks.

Workflows

  1. Earnings analysis: qveris_finance.event_calendar_earnings, qveris_finance.earnings_actual_surprise, qveris_finance.estimates_consensus, qveris_finance.fundamentals_is, qveris_finance.fundamentals_bs, qveris_finance.fundamentals_cf, qveris_finance.fundamentals_segment, qveris_finance.transcripts_earnings_call, qveris_finance.news_fin_tagged, qveris_finance.mkt_l1_rt.
  2. Peer comps: qveris_finance.ref_classification_industry, qveris_finance.ref_classification_theme, qveris_finance.fundamentals_derived_ratios, qveris_finance.mkt_l1_rt, qveris_finance.estimates_consensus.
  3. DCF/model update: qveris_finance.fundamentals_is, qveris_finance.fundamentals_bs, qveris_finance.fundamentals_cf, qveris_finance.estimates_consensus, qveris_finance.rates_govt_benchmark, qveris_finance.fx_spot, qveris_finance.mkt_l1_rt; output assumptions and sensitivity, not a target price promise.
  4. Market research: qveris_finance.research_analyst_reports, qveris_finance.news_fin_tagged, qveris_finance.event_calendar_corp, qveris_finance.ownership_institutional, qveris_finance.ownership_insider_trades; call qveris_finance.news_dedup_cluster only after cap-detail confirms it exists.

Live Fallback Policy

  • If qveris_finance.earnings_actual_surprise returns a provider error, fall back to qveris_finance.estimates_consensus plus qveris_finance.event_calendar_earnings; mark the earnings-surprise field as missing.
  • If qveris_finance.transcripts_earnings_call returns an error, use qveris_finance.news_fin_tagged only for context and do not invent management quotes.
  • If qveris_finance.fundamentals_cf or other statement payloads return a different fiscal period than the requested statement set, keep them out of aligned tables, retry once with stricter documented period parameters when available, and move unresolved mismatches to missing_fields plus data_quality.warnings.
  • Set qveris_trace[].fallback_used: true for fallback conclusions and include primary_tool_unavailable in missing_fields.
  • Do not state beat/miss as fact unless qveris_finance.earnings_actual_surprise succeeds or another QVeris CAP payload explicitly provides actual and consensus values.

Output Requirements

  • Return a Markdown user report by default, not a single large JSON object.
  • Use this report structure: Summary, Evidence Used, Analysis, Data Quality And Missing Fields, What This Can Support, What This Cannot Support, and Trace Appendix.
  • Put the user-facing conclusion before trace details. Keep internal controls, missing_fields, and data_quality readable in prose or compact tables.
  • Use a two-layer trace: concise user-facing evidence table by default, full qveris_trace JSON only in the appendix when useful, when the user asks for machine-readable output, or when preparing schema fixtures.
  • If max_calls, dry_run, or budget constraints prevent the main workflow from running, return a budget-limited Markdown report: state what was not called, do not infer missing facts, and list the next QVeris calls that would be needed.
  • Separate evidence, interpretation, uncertainty, and next verification steps.
  • End with a final non-empty line that is exactly Not investment advice.. Do not add a Chinese, bilingual, translated, or prefixed disclaimer line.

Prohibited Capabilities

Do not use original MCP/data-provider assumptions, any non-QVeris finance data provider, SEC scraping, browser automation, cookies, login state, third-party API keys, automated trading, wallet/swap, buy/sell points, portfolio action instructions, or target price commitments. Provider names are listed in the source record only for internal migration context; do not repeat them in final output.

References

  • Read references/qveris-tool-map.md before choosing tool calls.
  • Read ../references/qveris-finance-retry-policy.md when a CAP call fails, needs retry, or needs fallback classification.
  • Check ../references/qveris-finance-cap-registry-snapshot-2026-07-07.md when deciding whether a capability belongs on the primary path.
  • Use examples/default-markdown-report.md as the primary user-facing output example.
  • Use fixtures/qveris/sample-output.json, fixtures/qveris/fallback-output.json, and fixtures/qveris/budget-limited-output.json as schema fixtures only.
  • Use examples/natural-language-prompts.md for copyable natural-language test prompts.
  • Use examples/natural-language-test-output-2026-07-07.md as a dated reviewer output record.
  • Run scripts/validate_qveris_finance_report.py <markdown-report> on generated reviewer reports when updating examples or fixtures.
发现
标签

此技能尚未发布标签。

版本
最新版本元数据

版本

v2026.09.25

发布时间

2026年9月25日

分类

未分类

许可证

MIT

源路径

qveris-anthropic-financial-services

默认分支

main

最新提交

bb4e480

Tree SHA

adcc8d1