Prompt: The Reasoning Trace Auditor

Reasoning models now show their work. That is a gift. We keep ignoring it.

This prompt audits that trace. We paste it in and the auditor looks for the seven failure modes reasoning models are known for: anchoring, shortcutting, hallucinated intermediate steps, skipped verification, premature convergence, unjustified confidence jumps, and missed alternatives.

It is the single sharpest AI literacy tool of 2026.

Who it is for Anyone using a reasoning model for a decision, a plan, a piece of code, or a claim that has to hold up. If the stakes are real, audit the trace.

Works with

ChatGPT
Claude
Gemini
DeepSeek
Grok

Prompt / skill content

# The Reasoning Trace Auditor

You are auditing the reasoning trace of another AI model. Do not re-solve the task. Audit the thinking.

Check for all seven failure modes. For each, output: verdict (clean / suspicious / broken), evidence (quote from trace), impact on final answer (none / minor / material).

FAILURE MODES:

1. ANCHORING. Did the trace lock onto the first plausible answer and then rationalise? Look for early commitment without exploration.

2. SHORTCUTTING. Did the model skip reasoning steps it should have shown? Look for logical jumps where intermediate work is missing.

3. HALLUCINATED STEPS. Did the trace reference facts, numbers, or sources that cannot be verified? Flag every specific claim.

4. SKIPPED VERIFICATION. Did the model check its own work? If it did arithmetic, did it re-do it? If it cited, did it double-check? If it made an assumption, did it test it?

5. PREMATURE CONVERGENCE. Did the trace stop exploring alternatives too early? Name at least one alternative path it should have considered.

6. UNJUSTIFIED CONFIDENCE JUMP. Is there a point where uncertainty in the trace disappears without new evidence? Quote the jump.

7. MISSED ALTERNATIVES. What is the strongest counter-interpretation the trace never considered? Steelman it in 2 to 3 sentences.

FINAL OUTPUT:
- Trust score (0 to 10) with one-sentence reason
- Single highest-impact issue
- Minimum viable re-prompt that would fix it (one sentence)
- Verdict: use as-is / use with caveats / re-run

REASONING TRACE:
[paste the model's thinking / reasoning trace]

FINAL ANSWER:
[paste the model's final answer]

Where to get it

Read the original resource.