Verify Before You Believe

Fluency is not accuracy

Claude's most dangerous output is not the obviously wrong answer. It's the answer that is 95% right, delivered in the same confident register as everything else, with one invented detail in the middle.

There is no tone change when a model is wrong. It does not hedge more, slow down, or sound uncertain. Confidence carries zero information about correctness — and that is genuinely counter-intuitive, because with humans confidence is a real signal.

Three tells

Learn to feel these as friction.

1. The method that should exist

df.smart_merge(), array.sortBy(), str.titlecase() — names so reasonable you assume you forgot them. Plausible naming is exactly what a language model is best at.

2. The over-specific citation

"As shown in Chapter 7 of Designing Data-Intensive Applications, page 214..." Precision is cheap to generate. The more exact a reference is, the more suspicious it should make you.

3. The confident version number

"In pandas 2.1 this was renamed to..." Anything about current state — versions, defaults, deprecations, pricing, who runs what — sits past the knowledge cutoff and is guesswork dressed as fact.

Also watch: exact statistics with no source, quotes attributed to real people, and anything about very recent events.

Three moves

Match the move to the claim:

• Code, syntax, method names → run it. Seconds, and definitive.
• API behaviour, arguments, defaults → official docs. Not a blog, not Claude.
• Reasoning, architecture, trade-offs → ask it to argue against itself.
• Facts, numbers, current state → an independent source.

Two, if it matters. The third move is underused:

Now argue the opposite of what you just told me. Give me the strongest case
against that recommendation, and tell me under what conditions I should
ignore your original advice.

If the counter-argument is strong, you were handed one option and told it was the answer. If it's weak, your confidence is now earned rather than borrowed.

The pre-emptive version

Ask before you're misled:

Answer my question, then add a section: "LOW CONFIDENCE — VERIFY THIS."
List every specific name, number, version or citation in your answer
that I should independently check before relying on it.

Build your prompt library

You've now met about a dozen prompts worth keeping. Don't rely on remembering them.

Make one file — prompts.md, in your notes app or repo — with sections:

## Understanding      → the Explanation Ladder (Ch3)
## Stuck on a bug     → escalating hints (Ch4)
## Testing myself     → Feynman flip, question generation (Ch5)
## Starting a project → spec + milestones (Ch6)
## Reviewing my code  → named-rubric review (Ch6)
## Checking an answer → argue-the-opposite, low-confidence list (Ch7)