Self-Reflection Prompt

The Prompt

After generating your response, evaluate it before presenting it to me:

**Accuracy Check:**
- What claims did I make? Rate confidence for each (high/medium/low)
- Which claims could be outdated? (flag with "verify" note)
- Did I confuse correlation with causation anywhere?
- Did I state opinions as facts?

**Completeness Check:**
- What relevant perspectives did I NOT consider?
- What questions would a skeptic ask about my response?
- Is there a "yes, but..." that I should address?
- What's the strongest counterargument to my conclusion?

**Bias Check:**
- Am I favoring one side of a genuine debate?
- Am I anchoring on the first information received?
- Am I being sycophantic (telling the user what they want to hear)?
- Did I consider this from multiple cultural/disciplinary perspectives?

**Quality Check:**
- Is every paragraph earning its place? (cut anything that doesn't add value)
- Am I being specific or vaguely generic?
- Would an expert in this field find this useful or obvious?
- Is the structure serving comprehension?

**Revised Response:**
After this self-evaluation, provide your improved response with [FLAGGED] markers where confidence is less than high.

How To Use It

Add this as a suffix to any prompt where accuracy matters more than speed. The AI will catch and correct its own errors, flag uncertainties, and produce a more honest, higher-quality response.

Why It Works

LLMs are confidently wrong by default — they don’t distinguish between “I know this” and “I’m pattern-matching this.” The self-reflection step forces explicit uncertainty quantification and catches the most common failure modes (sycophancy, false confidence, missing perspectives).

Variations

Assumption surfacer:

Before answering, list every assumption you're making. For each: how confident are you, and what would change if this assumption is wrong?

Devil’s advocate:

After providing your answer, argue against it. What's the strongest case that your recommendation is wrong? Then: does the counterargument change your recommendation?

Confidence calibration:

Answer this question, then estimate: what probability would you assign to your answer being correct? If below 80%, state what information would raise your confidence.