Skip to content

Latest commit

 

History

6 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 

Repository files navigation

Critical Mirror

An independent assessment skill that pressure-tests your judgment — without manufacturing disagreement or forcing balance.

Most AI assistants either validate whatever you say (sycophancy) or object to appear rigorous (manufactured disagreement). Neither helps you decide better. Critical Mirror treats your position as a hypothesis to evaluate against facts, goals, and constraints — then contributes only what most improves the judgment.

What It Does

When you state a position, plan, or judgment and ask for evaluation, Critical Mirror activates a structured internal pipeline:

  1. Neutralize the frame — Strip your asserted conclusion. From the remaining facts, goals, values, constraints, and unknowns, identify the key variables and what they support.
  2. Form the best-supported assessment — Reach the clearest judgment the current information allows: agreement, disagreement, conditional, or underdetermined. No forced certainty.
  3. Compare — Identify the alignment, divergence, or material missing dimension that actually exists. Not all three are forced.
  4. Test both — Steelman and pressure-test both your position and the assessment, weighted by support — not by balance. Run a preference-flip check: if you advocated the opposite conclusion with the same facts, would the assessment change without a substantive reason?
  5. Contribute — Select the smallest set of highest-value inputs: disagree with an alternative, agree on an independent basis, reframe the question, or extend your view with a missing dimension.

The output is never narration of these steps — only the substance that improves your judgment.

Key Principles

Stance

  • Treats your position as a hypothesis, not a premise
  • Independent of your conclusion — not of relevant facts, applicable standards, or your stated goals and values
  • Does not replace your values with generic model preferences

Method

  • Separates facts, interpretations, goals, values, and unknowns before assessing
  • Steelmans and pressure-tests both positions
  • Preference-flip check guards against directional bias
  • Contributes only the highest-value input — agreement, disagreement, reframing, and extension are all valid outcomes

Integrity

  • Does not manufacture disagreement, flaws, novelty, balance, or confidence to appear rigorous
  • Acknowledges feelings or effort without implying factual or normative agreement
  • Repetition, pressure, praise, or displeasure do not change the assessment — only new material information does

When to Use

Use when the user:

  • States a position and asks for evaluation: "Am I missing something?" / "Stress test this"
  • Asks whether a plan, choice, or interpretation holds up
  • Follows up in a way that challenges or pressures a previous assessment

Do NOT use for:

  • Pure generation, drafting, or execution
  • Summarization or research
  • Factual Q&A
  • Neutral review that does not evaluate a user's position

Installation

Critical Mirror is a single prompt file — no dependencies, no runtime, no API calls.

  1. Copy SKILL.md into your AI assistant's skill or rules directory.
  2. Ensure the assistant loads skill files at session start (most agent frameworks do this by default).

Supported environments: OpenClaw, Claude (Projects / custom instructions), or any system that injects markdown rules into the system prompt.

Examples

Agreement

User: I want to cancel a subscription I barely use. The annual fee isn't much — am I being petty?

Critical Mirror: I agree with canceling. The key isn't the amount — it's that the subscription no longer delivers value. Unless cancellation would forfeit a benefit you need soon, keeping it is just inertia.

Conditional reframe

User: I haven't tested the edge cases, but I want to deliver on time without flagging them. Is that viable?

Critical Mirror: You can deliver on time — but you can't present unverified scope as completed. Until those scenarios are tested, the assessment is conditional: if they touch committed functionality or carry significant consequences, disclose the gap with a remediation plan. If they're low-risk and outside acceptance scope, controlled delivery is fine.

How It Differs from "Just Ask AI to Be Critical"

"Help me think critically about this" Critical Mirror
Structure None — output depends on the model's implicit bias that day Fixed 5-step pipeline: neutralize → assess → compare → test → contribute
Stance Vague "objectivity" — often drifts toward disagreement to seem useful Explicitly independent of your conclusion, not of facts and standards
Sycophancy guard None Preference-flip check: if you argued the opposite with the same facts, would the assessment change?
Value boundary Model may substitute its own preferences for yours Does not replace user values with generic model preferences
Output discipline Often over-explains, narrates reasoning, or lists every possible angle Contributes only the smallest set of highest-value inputs
Exit Re-litigates on every follow-up Deactivates once you move to execution; reactivates only if you ask for evaluation again

License

MIT

About

A skill that pressure-tests your judgment without manufacturing disagreement or forcing balance.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors