qverisai News sentiment radar
Standalone Execution Contract
Treat this skill folder as self-contained. When the skill is installed or copied alone, run commands from this directory and use scripts/run.mjs for dry-run, fixture, and live execution.
Do not hand-write QVeris curl or ad hoc API calls for normal operation. Manual QVeris calls are allowed only for debugging provider behavior, must be labelled manual_debug, and must not be reported as a successful skill E2E run. The skill E2E path is successful only when scripts/run.mjs produces the Markdown report, structured JSON, and trace artifact.
scripts/lib/qveris-runtime.mjs is bundled runtime plumbing for this skill package. No repository-level shared directory is required when using the skill as an installed package.
Natural-Language Invocation Contract
When this skill is triggered by a user request, treat the skill as responsible for the final artifacts. The user should not need to know or request a command. Produce these canonical outputs whenever the user asks for analysis, a report, or a reusable result:
- Markdown report
- Schema-valid business JSON
- QVeris trace JSON with tool IDs, providers, parameters, execution IDs, costs, skipped calls, and missing-data notes
Use scripts/run.mjs internally to produce the canonical outputs. Always pass a business JSON output path when producing artifacts. In the final response, link the report, business JSON, and trace, and summarize paid calls, credits, execution status, and missing-data limits.
Do not create alternate runners, alternate schemas, or one-off JSON shapes for normal use. If the canonical runner lacks a metric, state the gap in missing_data and improve this skill later; do not silently replace the skill with ad hoc code. Manual QVeris calls, web search, or provider-specific debugging may supplement the analysis only when labelled manual_debug; they cannot replace the canonical runner output or be reported as successful skill E2E.
If the user has not authorized paid QVeris calls, stop after dry-run/preflight or ask for approval. If the user authorizes QVeris spend, run live and stay within the stated budget.
Internal Deterministic Runner
Use the local runner internally before composing a free-form answer:
node scripts/run.mjs --dry-run --ticker NVDA --window-days 7
node scripts/run.mjs --live --ticker NVDA --window-days 7 --max-paid-calls 6 --max-credits 40 --output artifacts/live-smoke.md --json-output artifacts/live-smoke.json --trace artifacts/live-smoke-trace.json
The runner performs QVeris Discover / Inspect preflight, enforces paid-call and credit budgets, writes a Markdown report, writes schema-valid business JSON, and writes a JSON trace.
Workflow
- Clarify ticker, market, time window, user objective, and maximum paid QVeris Call budget.
- Use QVeris Discover (POST /search) to find the needed finance, market, filing, news, transcript, or social capabilities.
- Use QVeris Inspect (POST /tools/by-ids) before paid execution. Show coverage, parameters, latency, success signal, and billing_rule when available.
- Ask for explicit user approval before paid Call execution.
- Use QVeris Call (POST /tools/execute) only for the bounded sources needed for the task.
- Return the requested finance output with evidence strength, missing data, QVeris capabilities used, paid Call count, estimated credits, and a not-investment-advice disclaimer.
Output Contract
Return these sections unless the user asks for a narrower format:
- Objective and scope
- Data sources discovered and inspected
- Evidence table with source type, recency, and confidence
- Analysis or ranking
- Risks, dissenting evidence, and missing proof
- QVeris calls used and estimated credits
- Not investment advice
The JSON artifact must include catalyst_confidence_score, corroborating_roles, confirmation_roles, role-level missing_data, evidence roles, paid-call usage, and traceable execution IDs. A catalyst can be labelled confirmed_evidence_set only when news, aggregate sentiment, price reaction, and either filings or issuer/company-news confirmation return usable evidence.
Cost Guardrails
- Discover and Inspect are treated as free preflight actions.
- Paid actions are QVeris Call executions.
- Provider fallback attempts are also paid actions and must remain inside
--max-paid-callsand--max-credits; fallback attempts are recorded in the trace. - If estimated credits exceed the user's budget, reduce tickers, shorten windows, or ask for approval before continuing.
Methodology Reference
Read references/methodology.md when the user asks where the workflow comes from or how to adapt it. Read references/source-review.md for GitHub research and references/qveris-tool-map.md before changing QVeris data routing.
Source Inspiration
This skill is QVeris-native and does not copy source project text, prompts, code, or branding. It adapts public workflow patterns from permissively licensed finance AI projects:
- AI4Finance-Foundation/FinGPT (MIT): Financial news, social sentiment, retrieval-augmented financial text analysis, and forecasting input patterns.
- AI4Finance-Foundation/FinRobot (Apache-2.0): Equity research automation, report generation, multi-agent financial analysis, and risk assessment patterns.
- cooragent/ClarityFinance (Apache-2.0): Claude-skill style financial workflow, planning-with-files, multi-market coverage, and screening/dashboard patterns.