When Grok is the wrong tool

Grok told me, cheerfully, that a store policy let me return a kettle after 90 days if I was "reasonably dissatisfied." I laughed. Then I almost forwarded it. The store's page said 30 days, unopened, receipt, and they were not doing a bit.

That is the failure mode. Not evil. Just a good talker with a live search that still misses, plus a house style that makes the miss go down easy.

Step 3 is how you write in the thread when the facts are public and checkable. This one is about the jobs where you should close the tab, and the smaller wrongness you can catch if you stay.

Wrong tool, not bad mood

Grok is a public current-take machine that will also write and joke. It is a weak place to do deep work on a private codebase. It cannot see your repo unless you paste files, and pasting a repo is how you get a confident review of three files and silence about the other forty.

If the work lives in files, go to Cursor. I mean that in the boring sense. Open the folder. Let it see the project. Grok can talk about code you paste and it can be useful for a snippet or an error message you drop in. It is the wrong home for "why does checkout fail on staging and also what did we do in the payments folder last quarter."

Careful long legal-ish writing is the other miss. A landlord letter, a contractor agreement, a "does this clause mean they can fire me." Grok will give you fluent paragraphs that sound like a person who has been in a conference room. Fluency is the trap. I have seen it invent a notice window that felt standard and was not in the document I uploaded. For the long sit with messy notes and a rewrite you must stand behind, Claude is the better chair. Still not a lawyer. Just less of a stand-up act while it works.

Private company stuff belongs in a tool you chose on purpose, with your eyes open, not in a chat you opened because it was already a tab. I upload a PDF of a public policy. I do not upload the all-hands deck.

If you are still choosing tools at the problem level, which tool when is the comparison without the brand loyalty.

Funny and still wrong

xAI will tell you Grok answers with a touch of wit. You can push it further. That is entertaining on a slow night. It is a problem when the joke is doing the work of a source.

I asked about a product still being sold in 2026 and got a riff about how the company "quietly retired the thing people actually liked." Funny. The product page was still live. The riff was a vibe from old posts. I would have looked clever and been wrong in the same sentence.

Catch it like this. After any punchy answer, say "repeat the factual claims without jokes, each with a URL from this year." If the fun version had four claims and the plain version has two, you found the padding. The padding is where people get hurt.

Another tell: a specific number with no page. "$25 change fee" and a shrug. Ask where. If it cites a flyer from a travel blog, you do not have a fee. You have content.

Another tell: it agrees with you too fast. I once said "I think they killed that plan" and Grok piled on. They had not. I had given it a hypothesis and it helped me decorate it. Now I ask "argue the opposite for a paragraph, using sources I can open." If it cannot, my hypothesis was a mood.

And the citation that does not contain the quote. I click through. Ctrl-F the phrase. If it is not there, the model stitched a sentence it wanted. Tell it. "That quote is not on the page. Try again or drop the claim." It usually drops it. Sometimes it finds the real paragraph. Either way you just saved an email.

How I catch it in practice

I keep a dumb checklist on a sticky note, not because I am organized, because I forget when I am in a hurry.

Does this need to be true today, or just sound informed. If today, I need a date on the page.

Is the primary source a company or a government page, or is it a person on X. X is great for "people are mad." It is a weak place to hang a policy.

Did I paste the actual document, or did I describe it. Descriptions get invented clauses.

Would I be okay if this answer was a little wrong in a funny way. Return policy on a kettle, maybe I click first and live. Notice to a tenant, no.

I also ask Grok to grade its own certainty in plain words, not a fake percentage. "What part of this would you not bet a coffee on." The sentence it hedges is the one I was about to treat as done.

Usage limits can create a different kind of wrong. Free tier, you hit a wall, you rush the last question. Paid SuperGrok is a weekly pool across chat and the other toys, reset date lives under Settings, Usage. If you are slamming into a cap, stop and finish the check tomorrow. A half-read citation at 11pm is how the kettle story happens.

When I stay anyway

I stay in Grok when the question is public and moving. Outage chatter. Did this airline change the rule. Is the product on the homepage still the one from the 2025 reviews. Two options I can open in a browser.

I stay when I want a take I will argue with. I leave when I want a draft I will not argue with.

I leave when I notice I am asking it to remember a whole private system. "You know, the way we do refunds" and it does not know. It will guess a reasonable refunds process. Reasonable is not yours.

If you pasted code and it "fixed" it, read the fix like you do not trust a stranger in your repo. Because you dont. Run it. If you cannot run it, you were never coding, you were collecting comfort.

The recovery is boring. New chat, tighter prompt, primary URL pasted in by you. Or a different tool. Do not keep scolding the same thread for being the wrong shape. It will get more polite and stay the wrong shape.

Then if you are still here, and the job was the right shape, keep the thread on purpose. Step 5 is what to pin so next Tuesday you are not starting from a blank box and a fuzzy memory of a fee.

I still use Grok after it whiffs. I just do not let it be funny last.