How to use ChatGPT well

What ChatGPT is shaped for

The broad generalist. Widest range of everyday tasks, the largest ecosystem of extras, and a persistent preferences layer that most people never switch on.

• Wide general-purpose ability across writing, analysis, and code
• A persistent, editable memory of your preferences and context
• Projects that keep files and instructions scoped to one body of work
• Broad plugin and extension ecosystem

Where it gets its information

It answers from training plus a live web search when the question needs current information — and it does not always tell you clearly which it used. When currency matters, ask it to search and cite, and check whether the citations actually appear.

How to prompt it well

• Fill in custom instructions once. There is a persistent settings block describing who you are and how you want responses. Most people leave it empty and then repeat themselves in every chat. Put your role, your domain, and your format preferences there — it applies to every new conversation.
• Use Projects to separate contexts. A project keeps its own files and instructions, so your work chats stop leaking style and assumptions into your personal ones.
• Say "search the web for this" when you need current facts. Left to itself it may answer confidently from training data instead.
• Ask it to keep working in the same document rather than reprinting a whole draft each turn.

The fundamentals — goal, context, constraints, format — work the same everywhere. Build one in the prompt lab →

Giving it context

• Custom instructions are global; project instructions are scoped to that project; uploaded files live with the project, not the chat.
• Long chats drift — start a fresh one and paste the current state rather than fighting a sixty-turn thread.
• Tell it explicitly what not to assume; it fills gaps confidently rather than asking.

Getting the output format you want

• Ask for the container first ("a table with these four columns"), then the content.
• "Reply with only the JSON, no commentary" is followed reliably.
• Ask for a specific length in sentences, not words.

Files, documents, and research

Upload documents, spreadsheets, and images for analysis. Best results come from asking a specific question about a specific file rather than "summarise these." For anything numerical in a spreadsheet, ask it to show the steps it used, then spot-check one row yourself.

Example prompts

Rewrite this at a reading level a busy non-specialist can skim in 60 seconds. Keep the two numbers exactly as written. Give me three versions with different opening sentences.

Search the web for how this topic changed in the last six months. Give me five findings, each with the source link and its date. Flag anything you could not verify.

Here's the error and the function. Explain what's failing in two sentences, then give the corrected function only — no surrounding prose.

Teach me this concept in four steps. After each step, ask me one question and wait for my answer before continuing.

Using the attached contract, list every obligation that falls on us, with the clause number for each. If a clause is ambiguous, say so instead of interpreting it.

Give me 15 options, ordered from most conventional to most unusual. Do not evaluate them yet.

Turn this messy list of dates and names into a table sorted by date, and tell me which entries look incomplete.

Common mistakes

• Leaving custom instructions blank, then repeating the same preferences in every chat.
• Assuming it remembers a previous chat when memory settings vary by account and setting.
• Not noticing whether an answer came from training or a live search.
• Treating an eager, fluent answer as a checked answer.

What it will not do well

• Does not guarantee currency unless explicitly asked to search
• Does not reliably cite sources for claims made from training data
• Drifts from earlier constraints over a very long conversation

Privacy and what you send it

A consumer product with account-level controls over whether conversations may be used to improve models, plus separate business tiers with different data handling. Check which tier you are on before pasting anything confidential — the answer differs between a personal login and a work one.

How to check its work

• Ask for sources and open them; fluency is not evidence.
• Re-ask a factual claim in a fresh chat and see whether the answer holds.
• For anything arithmetic, ask for the steps and check one.