How to use the Research Bias Self-Check agent skill

What the skill makes the agent do

The skill is a checklist, not a research workflow. It names five thinking traps that often distort AI-written research: leader-bias (only looking at large companies), English-bias (missing Japanese, Korean, Taiwanese and European players whose sources are under-covered in English), narrative-bias (chasing concept labels like "AI stock" instead of the actual business), confirmation-bias (searching only for supporting evidence) and recency-bias (relying on an outdated figure because it ranks high).

For each bias the skill gives a correction. For example, the agent should search small and mid-caps and suppliers, check the bear case, and mark any figure older than a year as possibly stale. Before it writes conclusions, the agent is asked to confirm it cited at least one piece of disconfirming evidence.

• Reads the checklist at the start of the task, before the first search
• Broadens the query plan toward smaller companies, non-English markets, bear cases and recent data
• Re-checks the draft conclusions for missing contrary evidence and stale numbers

Install the skill

Download the ZIP file from the skill's page on this site. The package contains a single file, SKILL.md. You can install it in one of two ways.

For Claude Code, unzip the file into ~/.claude/skills/ to make it available in all your projects, or into a project's .claude/skills/ folder to limit it to that project.

In the Claude apps, open Settings, go to Capabilities, and upload the skill under Skills.

How to trigger the skill with natural requests

You do not need a special command. The skill is meant to run at the start of a research task, so ask for the task and mention that you want the bias check applied. The agent should then load the checklist before it begins searching.

Name the type of task clearly: a stock screen, a sector study or a company deep-dive. Say what you want the agent to avoid, such as only large caps or only bullish sources.

• Run the bias self-check, then screen Japanese and Taiwanese semiconductor equipment makers for a sector study.
• Before you write up a deep-dive on this company, apply the research-discipline checklist and find the bear case.
• Do a stock screen for AI infrastructure names and include small and mid-caps, not only the largest companies.

Examples of what changes in the output

Without the checklist, a screen for a supply-chain theme often returns the same handful of large companies. With it, the agent is asked to look for suppliers and smaller names and to say who is missing from the top ten that should be included.

In a company deep-dive, the agent should separate the actual product and financial statements from the label the company uses for itself. It should also present at least one piece of evidence against the thesis.

What you need and safety notes

The skill lists no requirements. It does not run scripts and does not use the network. It is a single markdown file that shapes how the agent reasons. The skill itself does not fetch or verify any data, so the agent's searches and any sources it uses are whatever your setup provides.

The skill pairs with two companion skills. financial_rigor (its cross_validate function) checks the numbers, report_audit checks the final report, and this skill checks the reasoning. Use them together for a layered review.

Tips for better results

Treat the checklist as a reset of attention, not a guarantee of accuracy. It improves coverage and honesty but does not make research correct.

Check the dates of any figure the agent gives you. The skill asks for this, but you should confirm it.

Ask the agent directly to name who it left out and what evidence argues against its conclusion.