| name | a-share-research-skill |
| description | Market-specific financial research skill for mainland China A-share analysis. Use when Codex needs to parse A-share filings, earnings reports, broker research, industry notes, announcements, market data, factor hypotheses, valuation work, catalyst tracking, or market-structure narratives with source-traced evidence and China-market conventions. |
A-share Research Skill
Operating Rule
Keep the market boundary first. Treat report parsing, factor engineering, and philosophical analysis as output modes inside A-share research, not separate top-level skills.
Preserve source-project capability before optimizing. When adapting a workflow, check references/source-map.md and references/capability-matrix.md so the result remains traceable to pinned upstream snapshots.
Output Modes
Choose one mode explicitly when the user has not specified the expected output:
report_parsing: extract, normalize, and explain content from filings, announcements, broker reports, earnings material, news, papers, or datasets.factor_engineering: turn structured research evidence into factor candidates with evidence, calculation logic, economic hypothesis, and validation status.philosophical_analysis: apply explicit thinking models to examine assumptions, causality, uncertainty, failure modes, and market narratives.
Workflow
- Identify the research market as A-share and reject crypto-style assumptions unless the user is comparing markets.
- Select the output mode and load only the relevant reference files.
- Run
scripts/source_doctor.py when live data sources are needed or when prior source health is unknown.
- Gather or inspect source evidence before forming conclusions.
- Run the Source Health Loop: detect failed, stale, partial, or suspicious sources and surface diagnostic events before conclusions.
- Apply the QuantMind Layer pattern: convert raw or semi-structured material into structured research evidence.
- If producing factor work, output factor candidates with the required factor contract.
- If applying thinking models, use progressive disclosure: start with the smallest relevant model subset and expand only when the question requires it.
- State data freshness, source limitations, fallback sources, repair candidates, and validation status.
Factor Candidate Contract
Every factor candidate must include:
factor_namemarketasset_universesource_evidencecalculation_logiceconomic_hypothesisvalidation_status
Do not present a candidate factor as a final trading signal without validation evidence.
References
references/source-map.md: pinned upstream source repositories and capability provenance.references/capability-matrix.md: source parity, market adaptation, and quality upgrade checklist.references/data-source-policy.md: A-share data-source tiers, fallback rules, and freshness handling.references/source-health-loop.md: diagnostic events, user warnings, fallback handling, and self-repair candidates for data-source problems.references/quantmind-layer.md: structured evidence extraction contract inspired by QuantMind.references/research-output-modes.md: output mode definitions and selection rules.references/thinking-model-adapter.md: progressive disclosure rules for thinking models.scripts/source_doctor.py: executable OpenCLI source health checker that emits diagnostic events and repair candidates.
