13 Claude Habits to Fix: Prompts for Agreeable, Confident, Forgetful Chats

Picture the first time Claude says something flatly wrong. It arrives in the same calm, well-organised paragraph used for everything else. There's no pause and no hedge, and nothing in the writing changes at all. The error only gets caught when the reader already knows the answer.

A lot of people start a note on their phone at that point. Things keep going slightly wrong, they assume the fault is their own, and they want to spot the pattern before another month disappears. After sixteen months of near-daily use, a note like that tends to settle at thirteen items.

None of the thirteen is about the AI being stupid. Each one is a specific behavior that nobody mentions, paired with a specific line you can type to stop it. The wrong answers were never the real trap. The trap was how easy the right-sounding ones were to accept.

The items are in the order people usually run into them, so feel free to jump to whichever one is happening right now.

The first week

1. The first answer is a draft

Ask what to do about something and a sensible-looking paragraph of general advice comes back, and the tab gets closed. The advice was general because Claude hasn't been told anything about you.

Instead of accepting reply one, type this:

Ask me three questions you need answered before you can give me
a useful answer to that.

Three questions about your real situation follow. Answer them, and the next reply is about you and not about everybody.

2. It agrees with you

Present a plan and Claude calls it solid. Present the very same plan as a friend's idea and ask for the flaws, and it finds four. Nothing about the plan changed, only how it was framed.

Give me the strongest case against what I just said. Do not
soften it.

Reach for this before any decision that costs money or can't be reversed.

3. It sounds identical when it's wrong

Nothing gives it away. Tone stays level, the sentence structure stays smooth, and an invented figure sits in the same steady sentence as a genuine one. The usual human cues for spotting uncertainty simply don't exist here.

Which parts of that are you least confident about, and what
would I check to verify them?

It will answer honestly. It just won't volunteer the information on its own.

4. Tidying beats ideas

Ask for an original idea and the output tends to be predictable. Give it forty scrappy notes and request a sorted table, and it does a very good job.

Keep the thinking for yourself. Pass along the sorting, the summarising and the first draft of whatever task has been put off.

The next few months

Once the early surprises fade, a second group of habits shows up. These take longer to notice because they don't produce an obvious failure.

5. Long conversations get worse

Two hours into a chat, answers turn vague and the details from the start are forgotten. That isn't the user's fault, and it isn't a bad day for the model.

The fix is a fresh chat with only the three things that actually matter pasted in. Sharpness returns within seconds. Almost everyone hits this, and almost nobody is told about it.

6. It won't say what it doesn't know

A question gets an answer because answering is the job. Claude fills a gap before it will name one.

Append this to anything that matters:

If you are not sure about any of this, say which part and why,
rather than guessing.

7. Dates, prices and opening hours are the weak spot

The risky zone is anything that moved lately: who holds an office today, a current price, whether a shop is open right now. Claude states those details with confidence while staying vague about whether they're still true.

The rule is simple. Anything containing a number or a date that you plan to act on gets checked at the source. Not asked again. Checked.

8. Yesterday is gone

A whole situation gets explained on Monday, and by Thursday Claude has never heard of you. Many people assume they did something wrong, or that it's ignoring them.

Store a short paragraph about yourself in a note: your role, current projects, and the style you want, such as brief and plain. Paste it at the start of any chat that needs context. A Project in Claude can hold it, so the pasting stops altogether.

9. Photograph instead of typing

People are usually most surprised by this tip. Typing out what's inside a folder, a drawer, a shelf of tins, or a stack of letters that keep getting avoided isn't needed. A photo does the job.

I am photographing these so I can search them later. Give me one
row per item: what it is, the key details, the date if there is
one, and anything I should act on. If a page will not read
properly, list it separately at the end rather than guessing.

Up to twenty images fit in one message. The final sentence of that prompt is the important part, because without it Claude invents rows for the pages it couldn't read.

The four that take a year

The last four tend to appear only after months of regular use, and they're the ones that cost the most when ignored.

10. Have it test you

A clean explanation slides in easily and creates a sense of understanding. Psychologists know this as the illusion of explanatory depth, and a paper from the ACM digital library suggests AI-written explanations strengthen it. The smoother the text, the more confident readers feel, and the weaker their results when they must use the idea.

Ease gets mistaken for evidence of learning. AI removes friction, and the lack of friction reads as mastery.

Now ask me three questions to check whether I actually
understood that.

Do this with anything you intend to rely on later.

11. Show your writing first

Asking for text in your voice with no samples produces a press release, because Claude has nothing to go on.

Here are three things I actually wrote. Study how I put
sentences together, then rewrite the draft below to match. Keep
my habits even where they are not the tidiest option.

Paste three real pieces: an email that was sent, a message, anything unedited. The difference shows up immediately.

12. Don't paste what you wouldn't read to a stranger on a train

According to a 2025 security study (covered by eSecurity Planet), 77% of users paste data into these tools. About one in five of those pastes held personal or payment information.

A few categories should never be pasted: bank details, passwords, medical information about other people, and anything with someone else's full name and address together.

When a document like that needs help, retype the relevant part without the identifying details. It takes a minute and removes the whole problem.

13. Decide what the saved time is for

The real gain here was about three hours a week, which is a modest number and not a dramatic one. Those hours didn't turn into rest. They filled with more of the same work, and months went by before anyone noticed.

Decide the purpose before starting, in a single written sentence, such as spending that hour in the garden. Without it, the hour flows to wherever unclaimed hours always go, and that is rarely a place anyone would pick.

Who can skip this

Anyone who has used these tools professionally for years can skip it. They found all thirteen long ago and have four more of their own.

Anyone who has decided this technology isn't for them, and is settled on that, can skip it too. Nothing here argues against that decision.

Developers hoping for Claude Code or engineering workflows can skip it as well. The context here is the chat window, used by somebody who doesn't write software for a living.

For someone starting from scratch

None of the thirteen is a reason to stop. Daily use continues, and so should yours.

Each one costs something only while it stays invisible. Once the pattern is known, the traps lose their grip: the first reply is a draft, it agrees with you, fluent doesn't mean true, and ease isn't learning. They become things you route around without thinking.

That covers everything. It doesn't warn against the technology. It records how quickly people stop double-checking a tool that has been correct eleven times running.

Which number is happening to you right now? A good guess is eight or thirteen, and thirteen is far more common than anyone admits.