What to do when something isn't working
In the last lesson we learned to ask well. In this lesson we'll learn what to do when the answer doesn't come out the way you wanted, which language is worth writing to Claude in, and we'll close with five ready prompts you can copy today.
If there's one lesson worth bookmarking, it's this one. The prompt toolbox at the end was built to be used again and again.
Habits 5 to 7: correct, clear, show
Habit 5: Correct early, and don't argue
- See Claude heading the wrong way? Stop it right away. Don't wait for it to finish. Better to stop in the middle, sharpen and continue. In Claude Code you stop it with the Esc key.
- Didn't come out the way you wanted? Rewrite the original request, instead of sending more and more small corrections.
- After two corrections that didn't work, open a new conversation. That's a rule from Claude Code's official docs: a clean conversation with a better request almost always beats a long conversation full of corrections. The long conversation "remembers" all the failed attempts, and they drag it down.
Habit 6: A new conversation for every topic
Don't run your whole business in one endless conversation. Claude's context is like a desk: when it's piled with papers from ten different tasks, the quality drops. Done with a topic? New conversation. And a new conversation doesn't mean starting from zero: if memory is on, Claude remembers what matters about you and your business.
Habit 7: Raw material, not descriptions
Instead of describing things to Claude, show them. This is an official Anthropic recommendation too:
- Instead of "the client sent me a slightly angry email", paste the email itself.
- Instead of "I want a design like that website", attach a screenshot.
- Instead of "I have a table with the numbers", attach the file.
Every description of yours loses details along the way. The original raw material doesn't.
And when is vague actually fine?
For exploring, vague is fine. "What would you improve on this sales page?" is an excellent question, because you really do want to hear directions. Claude Code's docs also say an open request can surface things you didn't think to ask. The rule: vague to explore, specific to execute.
Which language do you write in? An honest note
These lessons were first written in Hebrew, and plenty of people write to Claude in a language that isn't English. If that's you, or your audience reads another language, there are two facts worth knowing.
Not every language costs the same. Your limit is measured in tokens, small units of text, and Anthropic writes that the number of tokens for the same text also depends on the language. In Hebrew, for example, the same content breaks into more tokens, so it uses more of your limit.
In English, instructions are sometimes sharper. Anthropic measures performance in other languages relative to English, and notes that it varies from language to language. Hebrew doesn't appear in their table, so there's no official number for it. My experience: on complex tasks, instructions in English often bring a more precise result.
My recommendation, in three steps:
- Writing in English already? You're in the language Anthropic measures every other language against. Nothing to change.
- Need the output in another language? Keep the instructions in English and ask for the output in the language your audience reads. Say explicitly which language you want the answer in. That's an Anthropic recommendation too.
- English isn't the language you think in? Work in your own language and enjoy it. Everything in these lessons works great in Hebrew, where they were first written. Just write in your language's own alphabet, not transliterated into Latin letters, which is also an official recommendation. And a trick worth gold: instructions you write once and run many times, like the standing instructions in your settings or skills, are worth asking Claude to translate into precise English. You write them once in your language, and save on every run.
And what doesn't change in any language: your voice. What your audience reads stays in your own words.
The prompt toolbox: five ready to copy
Five well-built prompts, for five tasks that come up in almost every business. Notice how each one puts the habits to work: being specific, giving context, and defining "done". Replace whatever's in square brackets with your own details.
Price quote:
"Write a draft price quote for a client named [name], for [the service]. Scope: [short breakdown]. Price: [amount], [including or excluding] VAT. Structure: a short personal opening, what's included (up to 6 items), what's not included (2-3 items, to avoid misunderstandings), payment terms and how long the quote is valid. Tone: professional and warm, no apologies and no superlatives."
A gentle email to a client who's late paying:
"Write me a reminder email to a client whose invoice number [X] is [how long] overdue. Tone: friendly but clear, no apology and no threat. Include: the invoice number, the amount, and an invitation to get back to me if there's a problem. Length: up to 80 words. Give me two versions, one softer and one more direct."
From meeting notes to actions:
"Here's a transcript or notes from a meeting: [paste]. Give me: 1. A five-line summary. 2. A task table with the columns: task, owner, deadline. 3. A list of open issues that weren't decided. Anything that wasn't said explicitly in the meeting, don't infer it yourself: put it under 'to confirm'."
A business decision:
"I'm torn between [option A] and [option B]. The context: [a sentence or two about the business and the situation]. Build me a comparison table by cost, time, risk, and what happens a year from now. At the end, give me a reasoned recommendation, and also the strongest argument against your recommendation."
Content that sounds like you:
"Here are three texts I wrote myself: [paste]. Learn my style from them: sentences, rhythm, humor. Now write a draft post about [topic] in exactly the same voice. If you're not sure something sounds like me, put it in brackets."
The seven habits in a nutshell
- Specific beats vague.
- Context and a goal, not just an order.
- Define what "done" looks like, and let Claude check itself.
- Big task? Plan first. Huge? Interview first.
- Correct early. After two failed corrections, a new conversation.
- A new conversation for every topic.
- Real raw material beats descriptions.
Try it now
- Pick the prompt from the toolbox that's closest to a task that keeps coming back in your business.
- Send Claude:
"Here's a prompt I want to turn into a standard template for my business: [paste]. Ask me three questions about my business so you can tailor it, then give me back a tailored version, with square brackets only in the places that change from one time to the next."
- Try the new template on a real task.
- Save it somewhere easy to come back to. In stage 2 we'll learn to turn it into a standing instruction or a skill.
What to take from this
After two corrections that didn't work, don't argue: open a new conversation with a better request. And alongside that: a conversation per topic, raw material instead of descriptions, and a standard template for every task that repeats.
