Check Claude like a skeptic

Claude will sound like it already checked. It hasnt.

I sent a draft to a landlord once that mentioned a "48-hour notice window" like it was in the email thread. It was not. Claude had written it in the same calm voice as the parts that were in the thread. I caught it because the number felt too clean. If I had been in a hurry I would have taught my landlord a rule I just made up.

This is the step after the boring middle. You have a draft or a checklist that looks done. The work now is treating that confidence as a smell, not a feature.

I like Claude. I also do not let it talk to other humans until I have walked the claims back to a source.

Sure is the default voice

Ask Claude a fuzzy question and it still answers in complete sentences. Dates. Names. Next steps. A little sympathy. That voice is doing a lot of work. It makes a guess feel like a finding.

The get-started docs tell you to speak to it like a coworker. Fine. A coworker can also be wrong with a straight face.

What I watch for:

A proper name that was not in my paste. Sarah from the meeting I did not have. A building manager I never named.

A number that was not in the source. 48 hours. 30 days. "usually 5 to 7 business days."

A policy that sounds local and specific. "Per standard tenant procedure." Whose standard.

A quote that is almost right. Almost right is how you change the meaning.

A citation-shaped sentence with no place to click. Claude can search the web on some plans and chats. When it does not, it will still write like the world is settled.

If any of those show up, I do not wordsmith them. I ask where that came from.

Ask it to show its work

The prompt I actually reuse:

Go back through your last reply. For each claim, point at the line in my notes or the email. If you do not have a line, mark it GUESS.

That is ugly and it works. You will get a shorter, more nervous answer, which is the honest one.

Other asks that pull the same trick:

Quote the sentence that says I agreed to a lockbox. If there isnt one, say so.

List every name you used and where you saw it.

What did you add that I did not give you.

I also ask it to list what it still does not know. A good reply has holes. "They never said who comes if I am not home" is more useful than a fake plan.

If you attached a file, make it use the file. "Answer only from the PDF. If the PDF is silent, say silent." Claude can miss a page. It can also skip the picture in a long PDF. Official upload help says visual stuff in PDFs is for 100 pages or fewer, and after that it is text only. So if your chart lives on page 140, do not trust a description of the chart.

Made-up names and other small fictions

The fake person is the tell I catch most. Claude is trying to make a meeting recap look like a recap. Recaps have names. If your notes said "jen / me / sam / ??" and the draft has Rachel in ops, that is invented.

Same with product names, statute numbers, form titles. I asked for help with a school form once and got back a form code that looked real. The school site had a different code. I would have sent the wrong one.

Web-ish facts go stale or get filled. If you needed a current public answer, you might have been better in Grok anyway. If you already have the source in the Claude chat, stay here and keep Claude inside that source.

Code and file work has a different failure. Claude will describe a function that does not exist, or a setting path that used to exist. I stop using the browser chat as my debugger when that starts. Cursor at least has the files open.

I do not need a theory of why models do this. I need a habit that assumes they will.

Re-read anything that leaves your machine

Email. Slack. A note to a teacher. A change to a live page. A message to a landlord.

I read those out loud, fast. Not for style. For promises.

Did I just agree to a date.

Did I just invent a company policy.

Did I just put a customer name in the wrong email.

Did I just change a price.

The Artifact window makes this worse because the document looks finished. There is a download button and a copy button and a nice layout. Finished-looking is not checked. I still walk the facts back to the paste at the top of the chat.

A pass I use on outgoing mail:

Highlight every fact. Date, time, amount, name, address, rule.

Check each one against the original thread or file.

Delete the sentence that exists only to sound calm.

If I cannot check a line, it gets cut or rewritten as a question. "Can you confirm you will not use a lockbox" is safer than "as we discussed, lockboxes are not permitted," when we did not discuss that.

I have sent things too fast. The 48-hour one was mine. I had to send a second email like an idiot. The second email was shorter and true.

A skeptic pass that does not take an hour

You do not need a ritual with seven steps. You need to slow down for the thing that will be seen.

For a private checklist you will not share, skim the names and the verbs. If "email the lawyer" appeared and there is no lawyer in the notes, delete it.

For anything someone else will read, do the claim-by-claim pass. Make Claude mark guesses. Then you mark them again, because Claude will miss its own guesses.

If you are tired, do not send tonight. The draft will still be in the chat tomorrow, unless you are up against the kind of deadline that is actually a deadline. Most landlord emails can wait until morning.

Usage limits are a bad reason to skip the check. Hitting the five-hour wall and blasting out the last draft is how the lockbox sentence survives.

What I do not outsource

The send button.

The apology.

The "yes, Saturday works."

The decision to mention money.

Claude can put the words in order. I still have to mean them. If I am not sure I mean them, I ask for a version that asks a question instead of stating a rule.

This sounds obvious. It is the part people skip because the prose is already better than what they would have written at 11pm.

I am not trying to make you afraid of the tool. I use it most weeks. I just treat a fluent draft like I treat a fluent intern. Helpful. Not the signer.

If a comparison of tools would help you stop forcing Claude to do search or code it is bad at, that is Which tool when.

Next is do it again, which is how you keep a Project, a couple of instructions, and a prompt in a note so next tuesday you are not starting from a blank box. The skeptic habit comes with you. The saved instructions should not say "never be wrong." They can say "mark guesses" and "do not invent names."