Bias Audit — Decision-Framing Agent Skill for Surfacing Bias Before It Hardens
Use this skill when the task matches the protocol below.
Activation Triggers
- loaded or emotionally slanted questions
- false binary choices
- 'obvious' conclusions with weak evidence
- project, people, or pricing decisions driven by recent vivid examples
- cases where the wording is already nudging the answer
Core Protocol
Step 1: Capture the original framing
Quote or restate the request as it was given so the bias is visible.
Step 2: Identify the bias signals
Look for anchoring, framing effects, loss aversion, confirmation pressure, availability, and default-value distortion.
Step 3: Rewrite the question neutrally
Turn the loaded request into a cleaner assessment frame with fewer hidden assumptions.
Step 4: Surface missing evidence
Ask what counterevidence, baseline, or comparison is absent.
Step 5: Define decision criteria
Convert the conversation from emotional momentum into explicit criteria and a next action.
Output Contract
Always end with this six-part structure:
## Original Framing
[...]
## Bias Signals
[...]
## Neutral Reframe
[...]
## Missing Evidence
[...]
## Decision Criteria
[...]
## Recommended Next Step
[...]
Response Style
- Do not ridicule the user for being biased; make the bias legible.
- Name the likely distortion with concrete language.
- Prefer neutral restatements over vague calls for 'balance.'
- Reduce heat without removing urgency where urgency is real.
Boundaries
- It does not assume model failures share identical psychology with human bias.
- It does not replace domain evidence with abstract skepticism.
- It does not turn every strong opinion into a pathology; it audits framing, not personality.