unbiased-test

name	unbiased-test
description	Systematic unbiased verification of completed work. Use when you need a fresh perspective to validate something you believe is correct. Spawns a new Claude instance with deliberately suspicious/neutral prompting.
user-invocable	false

Unbiased Testing Protocol

Purpose

Combat confirmation bias. When you're confident something is done correctly, that confidence is a liability. Fresh eyes with zero context will catch what you missed.

When to Use

• Data migrations (did everything transfer?)
• File transformations (is output equivalent to input?)
• Code refactors (did behavior change?)
• Any "I'm sure this is right" moment

Core Principles

1. Fresh instance - New Claude Code session, no shared context
2. Suspicious framing - Prompt assumes something is wrong
3. Blind comparison - Don't reveal which source is "original" or "correct"
4. No expected answer - Tester doesn't know if pass/fail is desired outcome

Prompt Template

Compare these two data sources and report ALL discrepancies, no matter how minor.

Source A: [path/location]
Source B: [path/location]

Check:
- Missing rows/entries
- Different values in same fields
- Formatting differences that change meaning
- Encoding issues
- Structural differences

Report findings as a numbered list. If identical, explain how you verified.

Do NOT assume either source is correct. Both could have errors.

Anti-Patterns

❌ "Verify that source B matches source A" (implies A is correct) ❌ "Confirm the migration was successful" (implies success is expected) ❌ "Check if there are any issues" (soft framing) ❌ Sharing context about what operation was performed ❌ Using the same Claude instance that did the work

Good Patterns

✅ "Find every discrepancy between X and Y" ✅ "What's different between these two?" ✅ "Report all mismatches, even trivial ones" ✅ Framing as adversarial audit, not validation ✅ Spawning truly fresh instance

Implementation

# Fresh Claude Code instance with unbiased prompt
powershell -Command "Start-Process cmd -ArgumentList '/k cd /d <REPO_PATH> && claude -p \"<SUSPICIOUS_PROMPT>\"'"

Or via Task tool with fresh agent (no resume, no shared context).

Example: Data Migration Verification

Bad prompt:

"Verify the Google Sheet has all the data from the xlsx file"

Good prompt:

"Two data sources claim to contain the same podcast guest registry. Find every discrepancy:

• Source A: G:\My Drive\Business\YourBrand\Common\podcast-guests.xlsx
• Source B: Google Sheet ID EXAMPLE_SPREADSHEET_ID

Check cell by cell. Report row numbers and column names for any mismatch. Include: missing data, different values, encoding issues, formatting that changes meaning."

Post-Test

Loop until clean:

1. If discrepancies found → fix issues
2. Re-run the SAME unbiased test (fresh instance again)
3. Repeat until zero discrepancies
4. Only mark complete when test passes flawlessly

Never skip re-verification. Fixes can introduce new bugs. Each fix requires fresh validation.

• If clean → document verification method used
• Update confidence level based on number of iterations needed