Most people treat ChatGPT like a search engine that talks back. They type a question, skim the answer, close the tab. Fine for trivia, useless on a Tuesday afternoon with fourteen things on the list.
The people getting real leverage out of it do something different. They set it up once, feed it proper context, and run repeatable workflows through it. The gap between a casual user and a deliberate one is roughly the gap between owning a gym membership and owning a gym.
Here are the habits that actually move the needle.
1. Fill in custom instructions and never explain yourself again
There is a box in settings labelled custom instructions, and most accounts leave it empty. That is a waste. It is a one-time job that pays out in every single conversation afterwards.
What belongs in there
- Your role and the world you work in: “in-house marketer at a 40-person B2B software company” beats “I work in marketing”
- Default output shape. If you live in bullet points, say so. If you want tables, say so.
- Tone rules: plain English, contractions allowed, no corporate padding
- Pet peeves: no throat-clearing intros, no bolded headings for a two-line answer, no “delve”
- The facts you retype constantly: team size, product names, internal acronyms, the client who must never be named
Ten minutes of setup removes the warm-up sentence from every reply you get for the next year. It also stops the model from defaulting to generic consultant-speak.
2. Build a small prompt library instead of retyping
Any prompt you have typed more than twice belongs in a text file. Give each one a name you will recognise at 4pm on a Friday, something like STANDUP-CLEAN, EMAIL-DEESCALATE or BRIEF-FROM-TRANSCRIPT. Keep the version that worked, plus a one-line note about what made it work.
A strong prompt usually names four things: the audience, the format, the length, and what to leave out. Miss the last one and you get three paragraphs of context nobody asked for.
3. Paste the raw material, not your summary of it
If you want meeting notes cleaned up, don’t describe the meeting. Paste the messy 1,200-word transcript and ask for decisions, owners, deadlines and open questions in a table. The model is far better at distilling than at guessing.
Same rule for screenshots. ChatGPT reads images, so a photo of a whiteboard or a grab of a half-broken spreadsheet often beats three paragraphs of description. Describe the data, paste two example rows, state the output you want.
4. Teach it your voice with samples
Generic output is usually a context problem, not a model problem. Paste three things you actually wrote: a client email, a paragraph from a report, a Slack message you were pleased with. Then ask it to describe your style in ten bullet points before it writes anything.
Save that style sheet. Drop it into the prompt next time, or into a Project, and the first draft comes back close enough that you’re editing rather than rewriting.
5. Chain prompts into a workflow
One prompt rarely produces a finished piece. Four will.
- Brief: “Ask me six questions about the audience, angle and constraints. Don’t draft yet.”
- Outline: “Give me three structures, each with a different argument.”
- Draft in pieces: “Write section two only, 250 words, following the chosen outline.”
- Edit: “Act as a sceptical editor. Cut 20%, flag anything unsupported, then rewrite.”
Writing one section at a time beats asking for a 2,000-word draft in a single shot. Long drafts drift, and you end up editing the drift instead of the idea. The same logic scales up on the engineering side, which is part of how CyberAgent moves faster with ChatGPT Enterprise and Codex, turning specs into working code with an agent in the loop.
6. Use voice mode for the thinking part
Half of your best thinking happens away from a keyboard: on a walk, in the car, in the shower. Talk into the app for ninety seconds, let it be messy, then ask it to turn the ramble into five action items with owners and dates.
The same trick works for first drafts. Explain the argument out loud the way you’d explain it to a colleague, then ask for a structured version of what you just said. Spoken English is usually clearer than written English, because you can’t hide behind jargon when you’re talking.
7. Hand over the jobs you hate
When did you last write a regex by hand? Or fight a nested IF formula for twenty minutes? These are perfect delegation targets because they are small, checkable and boring.
- Spreadsheet formulas described in plain English, with two sample rows pasted in
- Regex, cron expressions and command-line flags you’d otherwise Google
- Cleaning a CSV: dedupe, fix date formats, split names into columns
- Turning a 40-message thread into a three-line status update
- Twenty subject line variations, then a shortlist of the three least annoying
Always paste an example of the input and the exact output format you expect. Then test it on one real case before trusting it on a hundred.
8. Ask it to argue with you
Before you send anything important, add this line: “You’re a sceptical CFO. Give me the three strongest objections to this proposal and tell me which part of my argument is weakest.”
It won’t replace a real reviewer. It will catch the obvious holes in thirty seconds, before a human has to point them out in front of other people.
9. Turn the wall of text into a picture
Plenty of ideas land better as a diagram than as six paragraphs. When you need something for a slide deck, paste your explanation into Napkin AI, which turns plain text into diagrams in seconds without touching a design tool. It saves the twenty minutes you’d otherwise spend dragging boxes around in PowerPoint.
10. Know when a different tool wins
If your whole day lives inside Word, Excel and Teams, you don’t need a separate tab. Copilot is already sitting in the sidebar, and for document work that proximity matters more than raw model quality. Our breakdown of Microsoft Copilot in 2025 covers where it genuinely helps and where it still fumbles, which should shape how much you lean on it.
The honest rule: the best tool is usually the one already open in front of you. Hopping between five assistants is its own form of procrastination.
11. Practise deliberately for a fortnight
Almost every tip above is really a lesson in specificity. Audience, format, constraints, examples. That’s a learnable skill, and it transfers to any model, which is why a couple of hours spent on free AI courses on Alison can pay off faster than another round of aimless experimenting.
Pick one habit from this list each week and use it until it’s automatic. Six weeks later you’ll have a working system rather than a bookmark, and you’ll notice the difference most on the days when the list is long and the patience is short.