How to use AI

Start with one task, not with a review of services#

The typical beginner mistake is a week spent reading comparisons and "best AI tools" lists. It is more useful to take a task you already do every week: make sense of a long email, rewrite a draft, turn a table into text, find the error in a spreadsheet formula. One familiar task gives you something no review can: the ability to tell a good result from a bad one, because you know the right answer yourself.

Which specific service you choose at the start barely matters. ChatGPT, Claude, Gemini, Copilot and the rest change every few months, and what is true of one today will be true of another tomorrow. Take whichever is available and works for you — and switch when you hit a limit.

What works well and what does not#

Task	How it goes
Rewrite, shorten, change the tone	Well, you see the result immediately
Explain an unfamiliar term or piece of code	Well, but check it on your own topic
Translate, proofread, break down an email	Well
Turn messy text into a table	Well, if the format is stated explicitly
Come up with options — headlines, ideas	Well, as a draft
Exact facts, figures, links, quotes	Badly: always needs checking
Arithmetic and counting letters in a word	Badly without a calculator or code
Recent events, prices, timetables	Badly, unless it has access to search
Anything about your personal data and files	Only what you have put in yourself

The rule is simple: a model is strong where you can judge the result and dangerous where you cannot.

Three ways to use a model#

A chat in the browser or an app. The simplest way in. Good for almost any one-off task.

Built-in assistants. A button in your document editor, your email, your code editor, your browser. Convenient because you do not have to copy text back and forth; the price is that you see less clearly what exactly was sent to the service.

An API — calling a model from a program. Needed when a task repeats hundreds of times: labelling reviews, sorting requests, producing summaries. This is where settings like temperature and the system prompt come in, along with tools and agent loops.

There is also the option of running local models on your own computer. They are usually weaker than cloud models and need decent hardware, but your data goes nowhere.

Habits that save time#

Put the material in the prompt. The model cannot see your screen, your drive or your inbox. Copy everything that matters in full; how to build a request from its parts is covered in how to write a prompt.

Start a new chat for a new task. An old conversation gets dragged along with the question and leaks into the answer.

Ask for a draft, not a final text. The attitude "this is a starting point I will finish myself" saves frustration and removes the temptation to send someone else's text as your own.

Ask it to show its work. "Show me which part of the document this conclusion comes from" makes checking three times faster.

Keep a collection of prompts that worked. A plain text file is enough: wording that worked once saves you time dozens of times.

Double-check what matters. The same question asked two different ways gives different answers — a disagreement is your signal to check the source.

What not to send, and the limits of all this#

Your data goes to someone else's servers. Passport details, medical records, bank statements, customers' personal data, work keys and passwords, documents under a non-disclosure agreement — none of that belongs in a chat. The service owner sets the terms for storing and using your conversations and can change them as they see fit.

How confident an answer sounds has nothing to do with whether it is right. The model produces the same smooth text when it is wrong and when it is right; asking it whether it is right is pointless — you just get another plausible text.

Check links, quotes, figures and names before you use them. This is the most common cause of embarrassment: citing a paper that does not exist is easy.

Health, law and money do not belong here. Looking up an unfamiliar term before a doctor's appointment is fine; making the decision is not. The same goes for a contract, a lawsuit, a savings account or a loan — those decisions are for a doctor, lawyer or financial adviser who is accountable for them.

Knowledge has a cut-off. Without search connected, a model does not know about recent events and can confidently invent details. It helps to understand the mechanics in advance: how neural networks work.

Step-by-step plan

1. Pick a task you do every weekOne where you can judge the quality of the result yourself.
2. Run it in one serviceNo comparisons or reviews: what matters is getting a first result and seeing the weak spots.
3. Check the answer against sourcesWrite out the figures, links and names and verify them; count how many mistakes you found.
4. Rewrite the promptAdd the material, the reader and the format — and compare with the old result.
5. Start a prompt collectionSave the wording that worked, together with the task it suits.
6. Set your own limitsWrite down which data you never send to a chat.

Check yourself

Sources

• OpenAI documentationHow calling models from a program works
  free
• Anthropic documentationRequest parameters, working with files and tools
  free
• Hugging FaceOpen models, including ones you can run yourself
  free

Was this helpful?

Yes No