Five habits that fix most bad Claude answers
The most common reason a Claude answer disappoints is not that the request was too hard. It is that the request left decisions unmade, and a model does not stop to ask the way a colleague would. It guesses, fluently, and you get something that is nobody's fault and no use.
Five habits remove almost all of that guessing. None of them is clever. All of them take seconds.
1. Say who it is for and what it is for
Compare "write about our new pricing" with this:
Write a short email to existing customers explaining that our pricing is changing on 1 November. They are small business owners, mostly non-technical. Plain and apologetic without grovelling. Around 150 words.
The second is not longer for the sake of it. Each clause removes a decision that would otherwise be made on your behalf: audience, medium, timing, tone, length. Those five are worth stating almost every time.
A test before you send: if a competent freelancer got this message and nothing else, would they come back with questions? Answer those now.
2. Give it the raw material
Claude cannot see your files, your inbox or your notes unless you hand them over. A request built on facts it does not have produces something confident and wrong.
Attach the document. Paste the numbers. Give it last quarter's version and say "match this format". The gap between "write a project update" and "write a project update from these notes, in the style of the attached example" is enormous, and closing it costs fifteen seconds.
This is also the cheapest defence against invention. Claude fabricates most readily in the gaps where it has nothing. Fill the gaps.
3. Ask for the format you actually want
If you want a table, ask for a table. If you want six bullets, say six.
And if you want a file, name the file. This is the single most under-used thing in the whole product. "Summarise this" gets you a message. "Summarise this as a Word document" gets you a real .docx with styles and headings. Ask for an Excel file and you get a workbook with live formulas rather than a picture of a table. Ask for a PowerPoint and you get an editable deck.
That routing happens because Anthropic ships skills for exactly this, and they sit in Customize > Skills where most people never look. The difference between a chat toy and something that produces work you can send is about six words.
One caveat worth knowing: for spreadsheets, add "every calculated cell must be a formula, not a hardcoded number". Without it you can get a workbook of static values that nobody can change.
4. Iterate instead of restarting
When an answer misses, the instinct is to rewrite the prompt and try again. Usually the better move is to stay in the thread and correct it, because everything already established stays in play.
Good, but too formal. Cut it by a third and lose the last paragraph.
That third point is wrong — we do not offer refunds after 30 days. Fix it and check the rest for the same mistake.
Each of those is faster than starting over. The exception is a thread that has genuinely gone off the rails, where earlier wrong turns keep pulling the answer back; then start fresh and carry the good parts with you.
5. Make it ask you questions
This is the habit almost nobody uses and the highest-leverage one on the list.
Before you write anything, ask me the five questions that would most change what you produce.
For anything substantial this beats a longer prompt every time. It surfaces the assumptions you did not know you were making, and it takes about a minute. It works because the hard part of most tasks is not the writing, it is deciding what the thing should be, and you are the only one who knows that.
The one that is not a habit
Verify anything that matters. Anthropic's own guidance is that Claude can produce responses that are incorrect or misleading, and that users should not rely on it as a single source of truth. Fabrication is most likely exactly where you are least able to check it — obscure specifics, precise figures, citations. A confident, specific, hard-to-verify claim is the highest-risk output there is, and polish is not evidence.
Two things help. Ask it to flag what it is unsure about, which works better than expected. And prefer giving over asking, which is habit 2 doing double duty.
All five at once: a worked example
Say the quarterly report is due and you have notes from four meetings and a spreadsheet of numbers.
Start by making it interview you, because you have not actually decided what this report is arguing:
Before drafting, ask me the five questions that would most change this report.
Then hand over the material and name the constraint that matters:
Draft this quarter's report following last quarter's structure exactly. Use the figures in the attached spreadsheet — do not estimate anything. Mark anything the notes do not cover as [TO CONFIRM].
That last instruction is doing the most work in this article. It converts invention into a visible gap, which turns checking the output from a search into a scan.
Then, and only then, ask for the artefact:
Now produce it as a Word document with the same heading styles as last quarter's.
Twenty minutes, most of it answering questions and checking figures. The pattern underneath — decide what to say, then make the thing — is behind most good results with Claude, and asking for both at once is behind most bad ones.
Where to go from here
Those five habits are the technique. The other half is the product: projects, memory, artifacts, connectors, incognito, and the skills that produce real files — most of it behind menus that are easy to never open.
The handbook covers all of it in 42 pages, with ten ordinary requests rewritten before and after, a field guide to the twelve moments people give up, four complete pieces of work start to finish, and a cheat sheet worth keeping in view. It comes with the Markdown source so you can paste sections into Claude and apply them to your own work, and with a small working skill that automates habit 5.