Red-Team Your A-Share Research: A Structured Challenge Skill

Where this fits

You've just finished a deep-dive on a Shanghai-listed semiconductor supplier. Your agent has assembled evidence, built a DCF, and the conclusion is bullish. Before you send it to the portfolio committee, you need a rigorous bear case—not a rubber-stamp "what could go wrong" list, but a structured challenge that forces your own reasoning to break. This skill automates that red-team process for A-share research, freezing your evidence and systematically probing for confirmation bias, hidden failure paths, and framework misuse.

Why agents benefit

• Freezes the evidence pack first – prevents the agent from retrofitting data to fit a counter-narrative; all challenges reference the original snapshot.
• Distinguishes structured challenge from independent review – enforces context separation, so you know whether the critique is truly blind or just a second pass.
• Builds falsifiable claims – rewrites your core conclusion into a testable sentence, exposing unfalsifiable statements as defects.
• Generates a failure-path matrix – maps causal chains, triggers, and time scales for each risk, without inventing arbitrary probabilities.
• Produces a minimal new-evidence plan – tells you exactly what data would discriminate between your thesis and the strongest alternative.

Practical scenarios

Pre-publication bear-case gate

Your research agent has a high-conviction "Buy" on a consumer staples company. Before the report goes out, you run this skill. It flags that the agent used current index membership to backfill historical peer multiples—a data integrity issue. The failure-path matrix reveals a regulatory risk the agent had dismissed as low-probability without evidence. You downgrade the conclusion to "Hold" and add two data points to the validation plan.

High-conviction thesis stress test

A portfolio manager asks you to double-check a thesis that has already survived two rounds of internal debate. You run the skill with a frozen snapshot. It identifies a survivorship bias in the comparable set and constructs a competing explanation: the company's growth is actually driven by one-time government subsidies, not structural demand. The skill notes the minimum discriminating data: quarterly subsidy disclosures.

Independent review for audit trail

Your compliance team requires an independent review of a major recommendation. You run the skill in a separate context with no access to the original reasoning chain. It performs its own evidence gathering and produces a challenge that is truly independent. The output includes the distinction between permanent loss and temporary volatility, helping the committee calibrate risk.

Add it to your agent workflow

Install the skill with:

npx skills add cyijun/china-financial-services --skill a-share-research-red-team

Then invoke it after your research agent has produced a final conclusion. Example agent instruction:

Run a-share-research-red-team on the current conclusion. Freeze the evidence pack and output the failure-path matrix and strongest alternative explanation.

Read the challenge cards in references/red-team-cards.md for structured prompts when you need to probe specific risk categories.

Compare before adopting

Before committing, compare this skill against:

• Manual red-team checklists – cheaper but inconsistent and biased by the same context.
• Generic adversarial-prompt skills – lack domain-specific A-share data integrity checks (e.g., future-row leakage, index membership backfill).
• Other financial research skills – check if they include the hard constraints: no fabricated mechanisms, no subjective probabilities, no trading advice.

Also verify maintenance freshness: the skill is tied to china-market-data conventions, so check recent commits for updates on data-source handling.

Why it is worth tracking

With 18 GitHub stars and a clear domain focus, this skill addresses a real gap: structured adversarial review for A-share research. It is not a generic "critique my essay" tool—it encodes domain-specific failure modes and data integrity checks. Watch for updates that expand the reference cards or add new bias checks. Evaluate it when you need to institutionalize red-teaming in your agent pipeline, especially before publishing high-conviction reports. For a broader catalog of agent skills, you can browse similar research and analysis skills at https://www.openagentskill.com/skills/cyijun-a-share-research-red-team.