Truth Prompt Setup and Verification Guide

The “Truth Prompt” is a third-party instruction set published by MavGPT. It was not developed by MIT, and a model's confidence score is not proof that an answer is true.

What this helps you do

Ask Claude to check its logic, call out uncertainty, and revise weak answers before responding. It is a useful quality-control layer, followed by real source verification when the stakes matter.

Before you start

You need a Claude account with access to profile instructions. The menu name can vary by plan and rollout. Open the MavGPT Truth Prompt guide and read the current version before copying it.

Terms to know

• Custom instructions: directions Claude applies to later conversations.
• Self-check: the model critiques its own draft. It is not an independent fact-check.
• Confidence score: the model's own estimate, not a measured probability.
• Primary source: the original research, product documentation, dataset, or official record behind a claim.

Step-by-step

1. In Claude, open your profile menu, then Settings.
2. Find the instructions or profile-preferences field available on your plan.
3. Copy the current Truth Prompt from the source page and paste it there.
4. Save the setting.
5. Ask a test question and request sources for every factual claim.
6. Open the cited sources and verify the important details yourself.

Copy this

Use this shorter, transparent version if you do not want to install the full third-party prompt:

Before answering, check your draft for unsupported facts, missing context, logical gaps, and one-sided framing.

In the final answer:
- Separate verified facts, reasonable inferences, and unknowns.
- Link primary sources for claims that could have changed.
- Do not invent a citation, quote, number, or source.
- If you cannot verify a claim, say so plainly.
- If the answer could materially affect health, law, money, safety, or privacy, tell me what needs independent verification.

A confidence score is optional. Never present it as measured accuracy.

Things to know

No prompt can force Claude or ChatGPT to “always tell the truth.” A model can repeat the same mistake during its self-check, cite a weak source, or assign high confidence to a wrong answer. The 0.8 threshold in the shared prompt is a rule for retrying, not evidence of 80% accuracy. Use primary sources or a qualified professional for high-stakes decisions.

If something goes wrong

• If the settings field is missing, keep the prompt in a Claude Project or paste it at the start of a chat.
• If citations do not open or support the claim, treat the claim as unverified.
• If Claude keeps retrying or becomes verbose, remove the numeric threshold and ask for one concise uncertainty check.
• If the prompt conflicts with a task, put the task-specific instruction in the current conversation and state which rule should win.

Final check

Ask a question with one deliberately missing fact. The setup passes only if Claude labels the gap instead of inventing an answer, and every cited source supports the sentence beside it.