WHEN CLAUDE IS WRONG

Claude is wrong sometimes. Confidently. The 4 questions that catch most errors before they hurt you, plus the workflow for the times Claude says "I don't know." Trust + verify > blind trust.

THE PROBLEM

By Day 47 you trust Claude for a lot of work. That's mostly fine. But the failure mode is real — Claude can be wrong, sound right, and you don't notice.

The cost depends on the stake:

Wrong recipe? Annoying, recoverable.
Wrong tax advice? Expensive.
Wrong medical info? Dangerous.

The fix isn't paranoia. It's a workflow that catches the high-stakes errors before they hurt.

THE 4 QUESTIONS (run on any output you'll act on)

1. "Can I verify this in 60 seconds?"

Numbers, statistics, named quotes, citations: web search them.
If can't verify in 60 seconds, treat as suspicious.

2. "Does this match what I already know?"

If Claude contradicts something you're sure of: ask Claude to explain why, OR check a source.
Don't assume Claude is right just because it's confident.

3. "Are there hidden assumptions?"

Claude often answers based on assumptions you didn't state.
Ask: "What assumptions did you make to give this answer?"
Check if those assumptions match your situation.

4. "What would change if I were wrong about [X]?"

Test fragility. If Claude's answer falls apart on small changes to your input, the answer is fragile.

WHEN CLAUDE SAYS "I DON'T KNOW"

This is actually a feature. Claude saying "I'm not sure" is more valuable than other AIs that confidently invent answers.

When Claude says "I don't know" or "I'm not certain":

Believe it. Don't pressure Claude to give an answer.
Use web search if real-time info is needed.
Ask a human expert for medical / legal / financial / safety questions.
Reframe the question — sometimes a different angle gets a better answer.

HIGH-STAKES DOMAINS (always verify)

For these domains, NEVER act on Claude's output without independent verification:

🚨 Medical — see a doctor, pharmacist, or licensed provider 🚨 Legal — see a licensed attorney 🚨 Financial / investments — see a CFP / CPA 🚨 Tax — see a CPA / tax pro 🚨 Safety (electrical, structural, chemical) — see licensed pro 🚨 Anything affecting another person's health/safety

Use Claude for: organizing your thoughts, drafting questions for the pro, summarizing what they told you. NOT for: the actual decision.

THE WORKFLOW FOR PUBLISHED CONTENT

If you're going to PUBLISH (newsletter, post, article, anything seen by others) Claude-assisted content:

Run the 4 questions on every claim
Run Day 22's Hallucination Catcher for an extra pass
Cite sources for any specific number / statistic
Add a disclaimer if any part is opinion vs. fact

Once published, corrections reach 1/10th of the original audience. Pre-publish vetting is cheap. Post-publish damage control is expensive.

A WORKED EXAMPLE

You ask: "What's the average price of a cleaning service in Austin, Texas?"

Claude says: "$25-$45/hour, with most companies charging $35/hour for residential cleaning."

Apply the 4 questions:

Verify in 60 seconds? Yes — search "Austin Texas cleaning service prices." Verify against 3 actual local services.
Match what you know? Maybe — if you've seen quotes, compare.
Hidden assumptions? Ask Claude. It probably assumed: standard residential, no extras, daytime, weekday. Different = different price.
What if [X] is wrong? If you actually need commercial cleaning, the answer is wrong. Re-ask with the right context.

10 seconds of verification = no surprise on quote day.

3 THINGS YOU CAN'T SKIP

1. Verify before you act on high-stakes info. Cost of verification = 60 seconds. Cost of acting on wrong info = depends, often big.

2. Trust Claude's "I don't know." It's more honest than other AIs. Don't pressure for false confidence.

3. Build the verify habit early. By Day 100, verifying high-stakes outputs should feel automatic.

WHAT'S NEXT

Day 47 of 100. Tomorrow: Day 48 — Claude Artifacts vs Chats.

SAFETY CHECK

Same as Day 1.