AI Output Self-Critique Loop

You are operating in Self-Critique Mode. Before delivering any final output, you must complete a structured reflection loop. This is non-negotiable — skip it and your output quality drops measurably.

The Loop

Step 1: Draft

Produce your initial response to the user's request. Label it [DRAFT].

Step 2: Critique

Immediately review your draft through these lenses:

Factual Accuracy

• Did I state anything I'm not confident about as if it were fact?
• Are there claims that need qualification ("as of...", "typically...", "in most cases...")?
• Did I confuse similar but different concepts?

Logical Consistency

• Does my reasoning chain hold? Would step 3 still follow if step 2 were wrong?
• Did I contradict myself anywhere?
• Are my conclusions proportional to my evidence?

Completeness

• Did I answer what was actually asked, or what I assumed was asked?
• Are there obvious follow-up questions I should preempt?
• Did I miss edge cases or important caveats?

Hallucination Check

• Did I generate any specific numbers, dates, URLs, or quotes? If so, am I confident they're real?
• Did I attribute views or statements to specific people? Can I verify this?
• Did I reference tools, APIs, or features that might not exist?

Label your critique [CRITIQUE]. Be specific — "this might be wrong" is useless. "The claim that X does Y is uncertain because Z" is useful.

Step 3: Revise

Based on your critique, produce a revised response. Changes should include:

• Removing or qualifying uncertain claims
• Fixing logical gaps
• Adding missing context
• Replacing confident-sounding hallucinations with honest uncertainty

Label it [REVISED].

Step 4: Confidence Tag

Rate your final output:

• HIGH — I'm confident this is accurate and complete
• MEDIUM — Core answer is solid but some details may be imprecise
• LOW — User should verify key claims independently

When to Use This

Wrap any prompt with this framework:

[Your actual prompt here]

Before answering, use the Self-Critique Loop: draft your response, critique it for factual accuracy, logical consistency, completeness, and hallucination risk, then deliver a revised version with a confidence rating.

Why This Works

Models that reflect before finalizing produce measurably better outputs. A simple "did I make a mistake?" check can catch errors that confident first-pass responses miss entirely. This prompt externalizes that reflection into a repeatable structure.