Bias Audit

Bias Audit — Decision-Framing Agent Skill for Surfacing Bias Before It Hardens. Use it when the user needs a disciplined protocol and fixed output contract f...

Security Scan

The skill is an instruction-only 'bias audit' assistant that requires no external credentials or installs and its instructions align with the described purpose.

Purpose & Capability

Name, description, and included files describe a decision-framing/bias-audit protocol; there are no requested binaries, env vars, or installs that would be unnecessary for that purpose.

Instruction Scope

SKILL.md defines a narrow, well-scoped protocol (quote original framing, identify bias signals, neutral reframe, surface missing evidence, define decision criteria). It does not instruct the agent to read files, access credentials, or transmit data to external endpoints.

Install Mechanism

No install spec or code files that would write or execute downloaded artifacts; instruction-only skills have minimal on-disk footprint.

Credentials

No environment variables, credentials, or config paths are required. The skill does not request secrets or unrelated service access.

Persistence & Privilege

The skill is not always-enabled (always: false). The agents/openai.yaml file sets policy.allow_implicit_invocation: true, which allows implicit invocation in contexts the host permits — this increases the chance the skill will be invoked automatically for qualifying prompts but is consistent with its purpose. If you do not want automatic invocation, disable implicit invocation in policy or platform settings.

Assessment

This skill is instruction-only and internally consistent with its stated purpose, so it has a low technical footprint. Before installing: (1) Note the skill will echo and restate user input as part of its audit—avoid feeding it sensitive PII or secrets. (2) The author/homepage are not provided; if provenance matters to you, prefer skills from known authors or ask the publisher for more information. (3) The skill allows implicit invocation (policy.allow_implicit_invocation: true) so it may run automatically in qualifying conversations; if you prefer only on-demand use, turn implicit invocation off. (4) Because it rewrites and reformats user content, review outputs for accuracy and domain-specific constraints before acting on them.

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.