A Practical Guide to Using ChatGPT and Gemini for Everyday Tasks

Most advice about using AI assistants stops at "you can use it to draft emails", which is true and useless. The gap between a mediocre result and a good one is almost never the tool — it is how much context you gave it and whether you asked for the right shape of answer.

What follows is the patterns I actually use, with the prompts written out, and an honest section on where these tools fail.

First, what these things are

ChatGPT (OpenAI) and Gemini (Google DeepMind) are both large language models. They predict text. Everything they do — summarising, drafting, explaining, writing code — is that one capability pointed at different problems.

This matters because it tells you what to expect. They are strong at anything where plausible and well-formed is most of the job: rephrasing, structuring, explaining, translating, brainstorming. They are weak wherever verifiably correct is the job, because nothing in how they work distinguishes a true statement from a fluent one.

Both have web and mobile apps, both handle images as input, and both offer paid tiers with more capable models. Features move constantly, so treat any specific claim about what they can do today — including mine — as worth checking.

Pattern 1: Give it the role, the audience, and the constraint

The single biggest improvement you can make is to stop writing one-line prompts.

Weak:

Write an email asking for a deadline extension.

Better:

You are writing as a freelance developer to a long-standing client.

Write an email asking to move a deadline from Friday to the following
Wednesday. The delay is my fault — a dependency upgrade broke the test
suite and took two days to unpick.

Be direct, take responsibility without grovelling, and propose what I
will deliver on Friday instead. Under 120 words. No subject line.

The second gives it a role, an audience, the actual situation, a tone, a length and an exclusion. Every one of those removes a decision it would otherwise make badly.

The length constraint matters more than people expect. Left unbounded, these models pad — that is what their training rewards.

Pattern 2: Ask for the structure you want

If you need a specific output shape, say so. Otherwise you get prose you then have to reformat.

Compare these three options as a markdown table with columns:
Option | Upfront cost | Ongoing cost | Main risk

Then a one-sentence recommendation. No preamble before the table.

"No preamble" earns its place. Left alone, these models open with a paragraph restating your question.

Pattern 3: Give it the source material

This is the difference between a tool that helps and one that invents things.

Instead of "what does the law say about X", paste the actual clause and ask it to explain that. Instead of "summarise the latest on Y", paste the article. Working from text you supplied means the answer can be checked against something concrete — and dramatically reduces the chance of a confident fabrication.

Here is our refund policy:

[paste]

A customer bought on 3 March and asked for a refund on 2 April.
Does the policy cover it? Quote the specific sentence you relied on.

That last line is the useful trick. Asking it to quote its source turns a vague answer into one you can verify in seconds — and if it cannot quote anything, you have learned the answer is not in the text.

Pattern 4: Iterate instead of restarting

When the first answer is close but wrong, most people rewrite the whole prompt. Respond to it instead:

Too formal, and the second paragraph repeats the first.
Cut it to three sentences and make it sound like a person.

The conversation carries the context. You are editing, not re-briefing.

Pattern 5: Make it ask you questions

For anything with a lot of unstated context, invert the usual direction:

I want to restructure how I organise my project files.
Before suggesting anything, ask me up to five questions about how I
currently work.

You get advice fitted to your situation rather than to the average of the internet — and the questions themselves often surface something you had not considered.

Tasks worth handing over

• Rewriting and shortening. Give it your own words and ask for tighter ones
• Explaining unfamiliar material. Paste the error, the contract, the log
• First drafts of formulaic writing. Release notes, meeting summaries
• Translation and tone adjustment, particularly into a second language
• Brainstorming, where volume matters more than precision
• Rubber-ducking, where explaining a problem to something that responds is the whole benefit

Tasks worth keeping

• Anything where a wrong fact is expensive. Medical, legal, financial
• Current events. Training data has a cutoff, and browsing is not the same as knowing
• Anything about specific people. Confident biographical invention is common and unkind
• Arithmetic you will rely on. They predict digits; they do not calculate
• Code you will not run. I have written about what breaks in AI-written articles, and the worst example was a code sample that mixed three incompatible API versions into something that looked idiomatic and raised an exception on line two

How to tell when you are being misled

Fluency is not evidence. These models are best at producing coherent text, which means a wrong answer arrives in exactly the same confident register as a right one. There is no tell in the writing.

Three habits that help:

1. Ask it to cite, then check the citation. Fabricated sources are usually obvious the moment you look for them.
2. Ask the same question in a fresh conversation. Genuine knowledge tends to be stable; invention often is not.
3. Watch for suspicious tidiness. Real subjects have exceptions and disputes. An answer with none has usually smoothed them away.

If you cannot verify a claim and it matters, treat it as unverified — not as probably fine because it sounded assured.

A note on privacy

Anything you paste may be retained, and on consumer tiers may be used to improve the service. Both providers offer ways to turn that off, and business tiers generally exclude it by default, but the settings move.

The practical rule: do not paste anything you would not email to a stranger. Client contracts, credentials, personal data about other people, unreleased work. For anything sensitive, redact the identifying parts first — the model rarely needs the real names to do the job.