Start With the Personal: How Everyday Tasks Teach You to Work With AI

You don't need a business case to start using artificial intelligence. Personal curiosity is enough. And the best training ground is everyday tasks you already know well: with them, you can see right away how AI reasons, where it makes mistakes, and how to phrase a request.

This article covers how to carry that experience over to work: where to look for work analogies, how to test safely, and what the official documentation recommends. It also includes ready-made prompts you can copy and adapt.

Familiar tasks are the best training ground

Imagine you're planning a vacation and ask AI to suggest a route. You know your budget, your dates, what you enjoy, and what you definitely don't want. So you'll spot a weak point in the answer immediately.

That's exactly what makes everyday tasks so useful for learning. They teach you three things:

• Seeing how AI reasons. You understand why it recommended this particular city or hotel and what data it relied on.
• Stating your criteria clearly. "Cheap and interesting" gets you a vague answer. "Up to 1,500 euros for two, 7 days, no overnight trips" gets you a specific one.
• Checking everything it generates. A train schedule, a ticket price, or museum opening hours may turn out to be outdated or made up. Everything AI produces is a draft, not a fact.

Here's what the difference looks like in practice. A vague request:

Suggest where to go on vacation.

A specific request:

There are two of us, planning a 7-day vacation in the second half of May.
Budget: up to 1,500 euros for two, not counting flights. We travel
by train, no overnight journeys. We love walkable cities and museums;
beach holidays don't interest us.

Suggest 3 routes. For each, make a table: cities, number of nights,
estimated accommodation and transport costs, pros and risks.
Separately, mark which details you can't guarantee and what I should
check myself.

That last sentence is the most useful one. It makes the AI point out the weak spots in its own answer.

Once the route is ready, ask the AI to critique its own response:

Check your plan for conflicts with my conditions: budget, dates,
no overnight journeys. List each violation separately. If something
can't be verified without up-to-date data, say so directly.

This doesn't replace your own check, but it often catches conditions that were missed.

The stakes here are low: you'll notice and fix a mistake in a vacation plan yourself, and the skill stays with you.

Look for analogies in actions, not topics

A vacation and your job rarely share a topic. But the actions you perform are often the same. So set the topic aside and ask yourself: what am I actually doing here?

"Choosing a vacation route" turns into a universal action: comparing options against given conditions and finding gaps in the data. And that is almost word for word a description of evaluating contractors.

Everyday task	Work task
Destinations or hotels	Contractors or suppliers
Budget, dates, convenient logistics	Cost, timelines, experience, contract terms
No information about airport transfers	No information about warranty obligations

The same prompt skeleton you used for the vacation works for a work task. Only the conditions change:

Compare three commercial proposals for an office renovation. The texts
are below; company names have been replaced with A, B and C.

Our conditions: budget up to UAH 400,000, completion by March 1,
a warranty of at least 12 months, payment in installments.

Make a table: cost, timeline, warranty, payment terms, what's included
in the price. Mark which proposal violates our conditions.
Separately, list what information is missing from each proposal and
what follow-up questions we should ask each contractor.

Pay attention to the last point. Here, the gaps in the data matter as much as the comparison itself: they become the plan for your next conversation with the contractor.

Try this exercise with a few of your own everyday scenarios. "Planning a weekly menu" is planning under constraints. "Writing a complaint to the property management company" is stating a demand clearly and backing it with arguments.

At work: carefully and gradually

A prompt that worked brilliantly for your vacation shouldn't go straight into a work project. Work tasks come with different constraints:

• Confidentiality. Client data, contract terms, and internal documents can't simply be pasted into a public chatbot. First find out which tools and data your company allows. Even then, replace names, people, and exact amounts with placeholders.
• Colleagues' expectations. Other people will read and use the result, so format and accuracy matter more.
• Higher standards. A mistake in a report or recommendation costs more than a poor choice of café.

So carry the skill over as a hypothesis to be tested. Start with small draft tasks: a document outline, a list of questions for a contractor, a comparison table built on anonymized data. If the result helps and checking it takes less time than doing the work from scratch, you can gradually expand how you use it.

Two examples of safe requests for a first test:

Draft the structure of an internal onboarding guide for a new
sales team member. Only sections and subsections, with one sentence
on what each covers, no finished text.

I'm preparing for a first meeting with a potential packaging supplier.
Write 15 questions grouped by topic: price, timelines, quality,
logistics, warranties. Mark 3 questions people often forget to ask.

Neither contains confidential data, and the result is easy to evaluate: you can see right away what to keep and what to discard.

A five-step algorithm

To carry a skill over from your personal life to work, take one of your own experiences with AI and answer five questions:

1. What result did I need? For example, three route options with an estimated budget.
2. What part did the AI do? It gathered options, compared them, and made a table.
3. What did I have to fix? Outdated prices, a museum closed for renovation, a forgotten date constraint.
4. Is there a similar action in my work? For example, comparing commercial proposals from contractors.
5. How will the conditions change when I carry it over? What data can't be shared, who will check the result, and what a mistake would cost.

If you can answer the fourth question and the fifth doesn't reveal any insurmountable obstacles, you have a candidate for your first work experiment.

AI itself can help with the fourth step:

I work as a procurement manager. In my personal life, AI has been
good at helping me with this action: comparing several options against
given conditions and finding what information is missing.

Name 5 typical tasks in my role that involve the same action.
For each, say what data it would need, which of that data might be
confidential, and which task would be the safest place to start.

Substitute your own role and your own action. The list you get isn't a finished solution but a set of candidates worth running through step five.

What AI developers recommend: seven techniques

Anthropic, the developer of the Claude models, publishes its own guide to writing prompts. Below are seven points from it, translated into the examples from this article. The principles are general, so they're worth trying with other chatbots too.

1. Write as if for a new colleague

The documentation suggests thinking of AI as a talented but new employee who doesn't know your rules or context. That gives you a simple test: show your prompt to a colleague who isn't familiar with the task. If they'd be confused, the AI will be too.

2. Explain why

The purpose behind an instruction gives the AI more to work with than the prohibition or requirement alone. Instead of "don't make it long," write:

I'll show this table to my manager at a 5-minute meeting, so keep only
what affects the choice of contractor.

Knowing the reason, the AI can work out for itself what matters and what can go.

3. Show examples

According to the documentation, examples are one of the most reliable ways to set the format, tone, and structure of a response. For the best results, it recommends 3–5 examples that are close to your task and varied enough. If all your examples are alike, the AI will copy some incidental feature they share. Say you're writing to a contractor: paste in two or three of your own successful emails, with names, people, and amounts removed first.

4. Separate instructions from materials

When a single request contains both a task and long texts, mark which is which. The documentation recommends tags, which reduce the risk of the AI mixing up parts of the request:

<proposal_A>
...text...
</proposal_A>

<proposal_B>
...text...
</proposal_B>

Compare the proposals against our conditions...

If tags feel unfamiliar, clear headings like "Proposal A:" will work too. What matters is that the AI can see where one document ends and the next begins.

5. Let it say "I don't know"

This is Anthropic's first tip against made-up facts: explicitly allow the AI to admit uncertainty. Add this to the end of your request:

If a proposal is missing information or you're not sure, write
"not enough information" instead of guessing.

6. Require a quote for every claim

When the AI is working with your documents, ask it to back up its conclusions with verbatim quotes. Anything that isn't in the text shouldn't make it into the answer:

Use only the texts of the proposals, not general knowledge. Add a
verbatim quote to every conclusion. If there's no quote, remove the
conclusion and put [] in its place.

This makes the answer easy to check: you compare the quote with the original instead of hunting for the source of a conclusion yourself.

7. Ask twice and compare

Run the same request several times in new chats. Discrepancies between the answers point to places where the AI may be making things up. You can also feed an answer back and ask the AI to check it, as in the route example above.

The documentation itself is candid: these techniques significantly reduce the number of errors, but they don't eliminate them entirely. Critical data still needs to be checked.

If no use case turns up, that's okay

Sometimes, after an honest analysis, you just don't see where AI would help in your current work. That's a result too, and a perfectly normal one.

No profession requires you to force a technology in where the usual way still works better for now. Your experience with the vacation, the menu, or the complaint to the property management company won't go anywhere: you already know how the tool works and where it can't be trusted. When the right task comes along, you'll be ready.

Sources

• Prompting best practices — Anthropic, Claude documentation
• Reduce hallucinations — Anthropic, Claude documentation