The truth nobody likes to hear
When someone tells me "I tried Claude, the results were so-so", I ask to see what they wrote to it. And again and again, it turns out the problem isn't Claude.
These tools are a mirror: a vague request gets a vague result. The good news? Talking to Claude well isn't an inborn talent or a six-month course. It's a small set of habits, and the first four are right here, in this lesson.
Everything you'll read here works just as well in the Claude app and in Claude Code. This is the lesson most worth coming back to.
Habit 1: Specific beats vague. Always.
This is the first general principle in Anthropic's official guide to writing requests for Claude. And the image they use is perfect: think of Claude as a brilliant new employee who doesn't know your procedures yet. The more exactly you explain what you want, the better the result.
| Not like this | Like this |
|---|---|
| "Sort out my files" | "Sort all the PDF files in the folder by date, into monthly folders named 2026-01, 2026-02 and so on" |
| "Write me a post" | "Write a Facebook post for an audience of self-employed people, about the most common pricing mistake. Up to 150 words, a direct and warm tone, no unnecessary exclamation marks" |
| "Check the report" | "Go through the attached report and check three things: totals that don't add up, duplicates, and rows with no date. Give me back a list of the problems only" |
Notice what's in the "Like this" column and missing from "Not like this": exactly what to do, exactly on what, and what a successful result looks like.
And here's a test worth adopting, also from Anthropic: imagine handing the request to a colleague who doesn't know the task. If they'd be confused, Claude will be too.
Habit 2: Give context, not just an order
Claude works much better when it understands what the work is for, not just what to do. Anthropic says this explicitly: when you explain to Claude why something matters, it understands the goal and aims for it. So instead of a rigid list of steps, tell it the story:
"The task: a price quote for a client who's torn between us and a cheaper competitor. The goal: show value without lowering the price. The audience: a small studio, so the tone is personal, not corporate."
And one more part of context, also in the official recommendations: give Claude a role. Open the request with "You're a business advisor who works with small businesses", or "You're a meticulous copy editor". According to Anthropic, even one sentence like that changes the behavior and the tone. The role guides all the small decisions along the way: the vocabulary, the level of detail, and what counts as a good answer.
Habit 3: Define what "done" looks like
The first item in Claude Code's official best practices is about exactly this: give Claude a way to check its own work. Without that check, "looks finished" is the only signal it has, and you become the one catching every mistake.
In your business, that translates simply:
- Asking for a table? Write exactly which columns.
- Asking for copy? Attach an example of a text you like: "Write in this style". According to Anthropic, examples are one of the most reliable ways to steer format, tone and structure.
- Asking for a calculation? Give it a checkpoint: "The total should come to X, and if it doesn't, flag it for me".
- Not sure about something? Ask: "Any item that isn't clear to you, mark it 'to check' instead of guessing".
That last sentence is your best protection against made-up answers. And one more tip from the same docs: ask Claude to show evidence, not just announce that it succeeded. "Show me exactly where this appears in the file" is worth far more than "I checked, everything's fine".
Habit 4: Big task? Plan first, then do
This is the habit that separates people who get good results from people who give up. Before a big task, ask:
"Before you start, show me a work plan. Don't do anything until I approve it."
Then you fix the plan, not the result. Cheaper, faster, and less frustrating. Claude Code has a dedicated mode for this, plan mode, and we'll get to it in stage 3.
And for truly big tasks, there's a trick Anthropic itself recommends:
"Before we start, interview me. Ask me every question you need to fully understand the task, and only then we'll build a plan."
Suddenly Claude is asking you exactly the questions you hadn't thought of.
So when do you skip the plan? When you can describe the whole task in one sentence. Claude Code's docs say it almost in those words: a small, clear change doesn't need a plan, just a good request.
Try it now
- Pick a real request you meant to send Claude this week. It can be the task you set aside in the first lesson.
- Open a new conversation and send:
"Here's a request I was going to send you: [paste]. Before you answer it, ask me up to five questions that will help you understand the goal, the audience and what a finished result looks like. Once I've answered, rewrite the request so it's precise, and only then do it."
- Answer its questions.
- Compare the rewritten request with your original. Whatever was missing from yours, write it in from the start from now on.
What to take from this
A vague request gets a vague result, and a precise request saves you rounds of fixes. Being specific, giving context, defining "done", and a plan before a big task: four habits that work from your very next request. Next lesson: three more habits, a word on which language to write in, and a toolbox of ready prompts.
