Most people trying to use ChatGPT effectively are working on the wrong thing. They’re hunting for better prompts while ignoring the machinery sitting one menu away — the settings that stop them having to explain themselves at the start of every single conversation.
If you open a blank chat, type a request, and paste in the same background context you pasted yesterday, you’re using maybe a tenth of what you’re paying for. And no amount of prompt cleverness fixes that.
These are the seven habits that actually change the output. Two of them take fifteen minutes once and then work forever.
Key Takeaways
- Custom instructions now hold 5,000 characters on paid plans, up from 1,500. Most people’s are years old or empty.
- Projects are the most underused feature — files and instructions that persist across every chat inside them.
- Upload the document. Describing what’s in a file to a tool that can read files is wasted effort.
- Correct the output specifically instead of regenerating. Starting over throws away context.
- Some things it still can’t do — and one popular feature was discontinued in April 2026.
Habit 1 — Write custom instructions properly, once
Settings → Personalization → Custom Instructions. This is standing context applied to every conversation you have.
The limit was raised in 2026 to 5,000 characters for Plus, Pro, Business, Enterprise and Education accounts — up from 1,500. That’s a meaningful change nobody announced loudly, and it means the cramped instructions you wrote in 2024 are now using a third of the space available.
What belongs in there:
- What you do, who you serve, what you sell — one short paragraph.
- How you want responses formatted. Length, bullets or prose, whether you want caveats.
- Words and habits you don’t want. Be specific: naming the phrases you hate works better than “avoid corporate jargon.”
- What to assume about your expertise, so it stops explaining things you already know.
Fifteen minutes, once. Then refine it over a fortnight as you notice what’s missing. Everything in writing better AI prompts applies here — you’re just writing the prompt that runs behind all your other prompts.
Habit 2 — Use Projects instead of endless new chats
The single most underused feature in the product.
A Project is a named workspace holding three things: files, its own instructions, and its own memory scope. Every chat inside it inherits all three. Set up a project for a client, a product line, or a recurring task, and you stop re-uploading and re-explaining.
Details worth knowing:
- Project instructions take precedence over your global custom instructions inside that project. One project can be a terse code reviewer while another is a long-form writer, with no global switching.
- File limits per project run around 5 on Free, 25 on Plus, and 40 on Pro, Business and Enterprise.
- You can now change a project’s memory setting after creating it — three-dot menu → Project settings → Memory. Project-only memory keeps that work sealed off: it won’t reference memories from outside, and its contents stay out of your other chats.
- Shared projects are always project-only and can’t be switched.
That project-only setting is the one to use for client work. It’s the closest thing to a clean separation between one client’s material and another’s.
Habit 3 — Upload the file, don’t describe it
It reads PDFs, Word documents, spreadsheets, CSVs, text, JSON, HTML, images, code and audio. Summarising a document in your own words so it can then summarise it back to you is a strange loop that a surprising number of people are stuck in.
This matters most for voice. Two examples of your own past writing, uploaded, will do more for output quality than any adjective you can think of. Showing beats describing, every time.
Habit 4 — Correct, don’t restart
When the first answer misses, most people delete it and rewrite the prompt from scratch. That throws away all the context that got it close.
Say what was wrong instead:
Second paragraph repeats the first. Cut to 120 words, open with the price, and drop the closing summary sentence — I never want those.
Specific correction beats a fresh attempt nearly every time. And when a correction is one you’d give repeatedly, move it up into your custom instructions so you never give it again.
Habit 5 — Make it interview you
The highest-leverage trick in this whole list, and almost nobody uses it.
Before you write anything, ask me the five questions you most need answered to do this well. Wait for my answers.
It works because the model knows what context it’s missing better than you know what context to supply. You get asked about the audience, the constraint, the thing you’d forgotten — and the output improves accordingly. Costs one extra turn.
Habit 6 — Use the right surface for the job
| Use | Good for | Not for |
|---|---|---|
| Quick chat | One-off questions, short drafts | Anything you’ll return to next week |
| Project | Ongoing work with reference files | A single throwaway question |
| Deep research | Multi-source reports with citations | Anything you need in two minutes |
| Canvas | Editing a document side by side | Quick answers |
| Voice | Thinking out loud, drafting while walking | Anything needing precision |
| Temporary chat | Sensitive one-offs you don’t want remembered | Work you’ll build on |
Most frustration in this product comes from using a quick chat for a job that needed a project, or waiting on deep research for something a normal answer would have handled.
Habit 7 — Keep a prompt library
When something produces output you’re happy with, save the prompt in a note file with a one-line label. Within a month you’ll have a dozen, and tasks that took thirty minutes of fiddling take two.
This is the mechanism behind almost every real productivity claim about AI — not the model, the reuse. It’s the same principle running through saving 10+ hours a week with AI and writing content faster with AI.
What still doesn’t work
It states wrong things confidently. Unchanged across every version. Statistics, dates, citations and prices arrive in the same assured tone whether they’re right or invented. Fact-checking AI content covers a routine that works.
Custom GPT creation is gone from personal accounts. As of 2026, new GPTs can only be created in Business, Enterprise and Edu workspaces — per OpenAI’s own documentation. Existing ones still work. Full detail in building a custom GPT for business.
Sora is gone. The video product was discontinued in April 2026. Any guide still telling you to generate video inside ChatGPT is out of date — a useful test for whether a tutorial has been maintained.
Arithmetic on your real numbers. Use a spreadsheet. Plausible arithmetic is worse than obviously wrong arithmetic.
Which model answered isn’t always obvious. Routing varies by tier and load, so the same prompt can land differently on different days. Our ChatGPT review goes through the tiers, and OpenAI’s pricing page has current details.
The bottom line
To use ChatGPT effectively, spend your first hour on setup rather than on prompts. Write real custom instructions. Build one project for the work you do most. Upload two examples of your own writing.
That hour changes every conversation you have afterwards. Better prompts only improve the one you’re currently typing.
Then add one habit a week from the list above. Trying all seven at once is how this becomes a productivity project instead of a productivity gain — and if you’re still deciding whether the paid tier is even worth it, start with free vs paid AI tools before you spend anything.
Frequently Asked Questions
How do I use ChatGPT effectively for work?
Set custom instructions once, create a Project for each ongoing area of work with its files attached, and save prompts that produce good output. Those three habits matter more than prompt wording.
What’s the difference between memory, custom instructions and projects?
Memory is automatic and global — ChatGPT decides what to save about you. Custom instructions are rules you write that apply everywhere. Projects are workspaces with their own files, instructions and memory scope, and their instructions take precedence inside them.
How long can custom instructions be?
Up to 5,000 characters on Plus, Pro, Business, Enterprise and Education plans, raised in 2026 from the previous 1,500. If you wrote yours before that, you have far more room than you’re using.
How many files can I add to a project?
Roughly 5 on Free, 25 on Plus, and 40 on Pro, Business and Enterprise. Files stay attached, so any chat inside the project can use them without re-uploading.
Can I stop a project mixing with my other chats?
Yes. Switch it to project-only memory via the three-dot menu and Project settings. It then won’t reference outside memories, and its contents stay out of your other conversations. Shared projects are always project-only.
Can I still create custom GPTs?
Not on personal accounts. As of 2026, creating and publishing new GPTs requires a Business, Enterprise or Edu workspace. Existing GPTs continue to work and can still be edited.
Can ChatGPT still generate video?
No. The Sora video product was discontinued in April 2026. Guides that still describe video generation inside ChatGPT haven’t been updated.
9 comments