Imagine: it's Tuesday morning, and you need to write to a customer whose job you had to push a week. You open Claude, type "write a polite email apologizing for the delay," and get back something that starts with "I hope this email finds you well" and reads like it could have come from any business on earth. So you close the tab, write it yourself in four minutes, wishing you hadn't wasted time with Claude.
But the issue wasn't Claude: you asked a question that could have been about anybody's customer, so you got back an answer that could have gone to anybody.
What's shifted
As the models have gotten smarter, so has their understanding of what you are asking. Which means we need to update our prompting to match. None of the old advice has stopped working, and Anthropic still recommends most of it for complicated jobs. But for the ordinary one-off tasks you'll use Claude for most days, these four carry less weight than the older guides suggest:
The expert costume. "Act as a world-class marketing strategist with 20 years of experience" improved results in 2023, when models needed the nudge to narrow their purview. Roles still help — Anthropic's own guidance says a role focuses Claude's tone and behavior, and that even a single sentence makes a difference. But the version they recommend is plain and functional ("You are a helpful coding assistant specializing in Python"). The world-class-with-20-years part was never the ingredient that worked. For a single task, telling it "you're writing for homeowners who just got three quotes and have no idea how to compare them" does far more than any job title you give it.
"Think step by step." Working through a problem before answering is now built into the models, without you telling it that explicitly. On a hard question you can still say "take your time with this one" and get something out of it, but as a magic phrase, it's not working as well as it used to.
Pressure tactics. Offering a tip, inventing a deadline, telling it your career depends on the answer — none of that was reliable back then and none of it does anything now.
The structured template. You know, the ones with a dozen labeled blocks (ROLE, CONTEXT, CONSTRAINTS, TONE, OUTPUT FORMAT, "do not hallucinate," "take a deep breath") wrapped around one actual request. Labeled sections do help on a complicated prompt: when you're handing over instructions, a document, examples, and your own numbers all at once, marking which is which keeps them from running together, and Anthropic recommends doing it. For "write a reply to this customer," it's overhead. Just make the request, and add the facts.
What all four have in common is that none of them tell the model anything about the context. Claude doesn’t know what’s in your inbox, customer records, or files unless you paste it, upload it, or give Claude access to it. Every useful thing you can add to a prompt is a fact about your situation that it has no other way to get.
Ok, enough of what matters less than it used to.
Here's what actually works to improve your results
Six habits, in rough order of how much they improve what comes back.
1. Give it the real material
Instead of "write a reply to a customer who's annoyed about a delay," paste the customer's actual email, and some details that explain the delay. Instead of "summarize my sales last quarter," attach the spreadsheet. Instead of "write a bio in my voice," upload three things you've already written (and do this in a project, so it's repeatable).
This is the single biggest quality jump available to you, and it's less work than writing a good description — the material already exists.
2. Say who it's for and where it's going
"Write a rescheduling message in a friendly and engaging tone" and "this is a text to a client I've already rescheduled once and want to keep for the next five years" produce completely different writing. Audience and destination do most of the work that tone instructions are trying to do.
3. Show one thing you liked
If you want it to sound like you, paste in one email you were happy with and say "match this voice." One real example beats three paragraphs describing your tone, because your tone is made of specifics you probably can't articulate.
This has a name in AI circles — few-shot prompting — but it just means showing examples. Two or three is plenty. You don't need a library. (Anthropic's own prompt engineering guide says the same thing, if you want the technical version.)
4. Ask it to interview you first
For anything longer than an email, this is a great move:
Before you write anything, ask me up to five questions you'd need answered to do this well.
You'll get asked about things you'd never have thought to include — who's actually reading this, what happened the last time, what you can't promise. Two minutes of answering turns a guess into a briefing. It's also the fastest way to find out that you weren't sure what you wanted, which is much better to discover now than after getting frustrated because you're not getting the quality you need.
5. Say what "done" looks like, and what to avoid
Tell it how long, what format, and what to leave out. "Under 200 words, no bullet points, don't mention pricing" removes a whole round of edits before it happens.
Banning specific habits works better than describing the ideal. "Don't use 'I hope this finds you well,' don't open with a compliment, never use the words leverage or seamless" beats "write naturally," because "naturally" means nothing to a model and "don't say seamless" is unambiguous.
6. Correct inside the conversation instead of starting over
When the first draft misses, the instinct is to close it and go write a better prompt, and that's usually the wrong move. "Shorter, cut the last paragraph, and the second line still sounds like a brochure" gets you there faster, because everything already in the conversation still counts — the file you pasted, the example, the three things you already fixed.
And when you notice you're making the same correction for the third time, that correction needs to be saved. Move it up into your standing Instructions for Claude (This post covers where that lives) so every future conversation starts with it already handled.
A before and after
Before:
Write a blog post about spring HVAC maintenance.
After:
I run a residential HVAC company in Longmont, CO — mostly older homes, mostly repeat customers. Write a blog post for homeowners about what to do before they run the AC for the first time this spring.
They're handy enough to change a filter but won't touch anything electrical, so don't assume they know what a condenser is. Around 700 words, and lead with what they should do this week.
Include the three things we get called about every May: filters nobody changed all winter, outdoor units packed with leaves, and thermostats that were never switched over. Say plainly which of those they can handle themselves and which needs us.
Don't open with "as the weather warms up." Here's a post I wrote last year whose tone I liked: [paste].
The second one took a few minutes to write (even less if you're dictating!), but it's the difference between a draft you throw away and a draft you edit for ten minutes and publish. It's longer, and every line of that length is a fact about the business the model had no other way to get.
Of course, this applies outside writing, too. Instead of “help me choose between three supplier quotes,” attach the quotes and say: “Compare these for total cost, lead time, warranty, payment terms, and anything one includes that the others don’t. Put the differences in a table, flag anything that isn’t clear, and don’t choose a winner unless the documents support it.” Now Claude has the actual options, your definition of “best,” and a clear job to do.
One more thing: check the work
So, the part that still requires you: models still make things up, and they're fluent enough now that a wrong answer sounds a lot like a right one. A few habits to keep:
- When facts matter, ask for sources and then check the links to make sure they're legit.
- Ask "what assumptions have you made here?"
- Challenge it when something seems off.
- Drop the response into a different LLM and ask it to critique it.
It stress tests the results, and bottom line, you're still responsible for the end result.
Where this leaves you
Every one of these is what you'd already do briefing a new hire in their first week: hand them the real files, tell them who they're writing for, show them one thing you liked, let them ask questions, be clear about what finished looks like, and correct them as you go. Treat it just the same here.
The people getting real work out of AI are the ones who give it enough to work with. It costs you the two minutes it takes to say what you already know, but it'll save you rounds of edits, or worse yet, abandoning the results to just write your own.
Related: Claude, Cowork, and Code — what's the difference? · How to set up Claude as your co-worker
Want help figuring out where AI actually fits in your business? Take the free AI Assessment — a few questions about how you work, and you'll get your stage on the AI path plus the practical next steps to move up one.