25 ChatGPT Tips and Tricks You Probably Did Not Know in 2026

There is a category error in most ChatGPT advice, and it costs people more than any prompting mistake.

The advice concentrates almost entirely on what you type into the box. Meanwhile the things that most change your results sit in settings nobody opens, in features that persist across conversations, and in the difference between pasting text and uploading the actual file.

Prompt wording matters less than it used to, because models infer intent better every release. Configuration matters more than it used to, because there is now considerably more of it. This is about the second category.

Set Up Custom Instructions Before Anything Else

If you do one thing from this article, this is it, and it takes ten minutes.

Custom instructions apply to every conversation automatically. Rather than explaining your job, your industry, and your preferred style each time, you state it once and it is always in effect.

What actually earns its place in there is more specific than most people write. Your role and sector, so it stops explaining your own field back to you. Who you typically write for. And concrete style constraints stated as prohibitions — "do not open by restating my question", "no bullet lists unless I ask", "say when you are uncertain rather than guessing". That last instruction changes output quality more than any prompting technique I know of.

What does not belong: vague aspirations like "be helpful" or "give great answers". These do nothing. Constraints work; encouragement does not. The custom instructions guide goes deeper on what to put where.

Memory: Useful, And Worth Pruning

Memory carries details between conversations rather than starting fresh each time. When it works, it is genuinely pleasant — it knows what you do, remembers your preferences, stops asking the same clarifying questions.

The failure mode is accumulation. It also retains things from one-off tasks, and those persist. Help a friend draft something in an unfamiliar register once and it may quietly assume that register is a preference of yours for weeks.

Open the memory settings occasionally and read what is actually stored. Delete anything that was situational. This takes two minutes and prevents the slow drift where responses become subtly wrong for reasons you cannot identify — which is a genuinely annoying problem to debug from the output alone.

You can also tell it directly to remember something, which is more reliable than hoping it infers what mattered.

Projects, When The Work Is Ongoing

Projects group related conversations and files with their own persistent instructions attached.

The distinction that matters: use them when you return to the same body of work repeatedly. A book, a client, a research area, an ongoing analysis. The project holds the context so you are not rebuilding it every session, and its instructions can be far more specific than your global ones.

For one-off questions they are pure overhead. The people who get most from them are those with a handful of long-running threads rather than many unrelated queries.

Upload The File, Do Not Paste The Text

This is the most commonly missed practical upgrade.

People paste an extract of a document into the chat and ask about it. Uploading the actual file is meaningfully better — the model works from the full document with its structure intact rather than from whatever fragment you selected, which means it can catch things you did not know to extract.

It handles spreadsheets and data files properly too, running actual analysis rather than pattern-matching over numbers in text. If you have ever pasted a table and received confidently wrong arithmetic, this is the fix.

For working on a document rather than asking questions about one, Canvas is the better surface — it keeps a single draft you both edit instead of scattering revisions up the conversation.

The obvious caveat applies and applies harder here than elsewhere: uploading a file hands over the whole file. Client material, personal data, and anything under NDA should not go in, and what happens to your data covers the specifics. Anonymise first — replace names with roles and real figures with representative ones — and the analysis is just as useful.

Two Habits Worth More Than Any Trick

Make it argue against itself. After it produces something you are reasonably happy with, ask for the strongest objection to it. This is the single most useful follow-up available, because the default output is agreeable — it is built to be — and agreeable output is exactly what fails to catch the flaw in your reasoning. Ask what a well-informed critic would say and the response is frequently more valuable than the original.

Iterate instead of perfecting. Send something adequate, read it, correct it. "Shorter." "You have missed the main objection." "Cut the introduction." Three rounds of that in under a minute beats a carefully constructed opening prompt, because each correction responds to something real rather than to a guess. More on this in writing better prompts.

Free Versus Paid, Honestly

The free tier is genuinely capable for the four things ChatGPT is actually good at — drafting, restructuring, generating options, and explaining. Most people asking whether to upgrade do not need to yet.

Paying makes sense when you hit usage limits often enough to disrupt your work, when you do heavy file or data analysis, or when you need the stronger reasoning models for genuinely difficult problems rather than routine drafting.

The advice I would give: use the free tier until it visibly obstructs you, then upgrade. Subscribing in anticipation of needs you have not encountered is how people end up paying for tools they use twice a month.

What Not To Bother With

Not every feature earns attention. Voice mode is impressive and, for most desk work, slower than typing. Image generation inside a chat is fine for casual use and outclassed by dedicated tools if the output matters — the image generation guide covers those. And the extended reasoning modes are genuinely better on hard problems while being slower and unnecessary for the routine work that makes up most usage.

Being selective is the point. The feature list grows every quarter, most additions are not for you, and treating each one as something you ought to adopt is a reliable way to spend more time managing the tool than using it.

For applying this in a professional setting, how to use ChatGPT for work covers where the output is trustworthy and where it is not. If you are weighing it against other assistants, ChatGPT vs Claude and ChatGPT alternatives cover the wider field.