Setting up ChatGPT properly takes about 20 minutes and makes every conversation after that significantly more useful. The key is using Projects, custom instructions, and a couple of reference files to give it context that sticks – rather than starting from scratch every time. This guide walks you through exactly how to do that, including a ready-made instructions block you can copy straight into your account.
1. Stop treating ChatGPT like Google with a chat box
If you keep opening a fresh chat, pasting a long prompt and hoping for magic, you’ll keep getting generic waffle. ChatGPT needs context, constraints, and clear examples before it can give you anything that doesn’t sound like LinkedIn on autopilot.
OpenAI’s Projects feature is built exactly for this: a single space where you set the goal, upload your examples, and give it standing instructions. Think of it like hiring a part-time member of staff, giving them a proper induction, then keeping all their work in one place – instead of sacking them at the end of every shift.
2. Before you type anything
When you open ChatGPT, don’t just click “New chat” and start rambling. Take 60 seconds to set the ground rules:
Turn on extended thinking for serious tasks so it actually shows its working and checks its logic – instead of blurting the first idea that comes to mind.
Turn on web search when you’re dealing with anything time-sensitive like SEO, regulations, or trends, so it pulls in up-to-date information instead of guessing.
Decide what this workspace is for, in plain language, before you ask the first question. For example: “This is for planning and writing weekly blog posts for our small business clients.”
The big mindset shift: AI can’t read your mind. If you haven’t said it, it doesn’t know it – and it will absolutely bluff to fill the gaps.
3. Create one proper Project
Instead of random chats, create a single Project for each repeat job: content, emails, proposals, SOPs, that sort of thing. Set it up once, reuse it every week.
For a small agency or local business, some good starting points:
“28 Pixels blog content engine” (or swap in your own business name)
“[Your Business] website copy and updates”
“Client proposal drafting”
In each Project, do three things.
3.1 Name the job, not the tool
Call it what it actually does in your world. “Monthly newsletter writer for our clients” is clearer than “Marketing AI”. That name will remind you what context lives there – and stop you stuffing everything into one bloated project trying to do 47 jobs badly.
3.2 Add 1-3 reference files
Upload a few pieces of writing you’re genuinely happy to put your name on. For example:
A blog post that sounds like you. A services page that converts enquiries. Any quick brand or tone notes you already give writers.
ChatGPT reads these and picks up your rhythm, phrasing, and level of detail. It’s not perfect, but it’s a lot closer than starting from a blank slate and praying it just somehow “gets you”.
3.3 Set project-specific instructions
Each Project gets its own instructions, separate from your global account custom instructions. This is where you tell it how to behave for this one job.
For a content project, you might say:
“You help me plan and write SEO-friendly blog posts and web copy for small, service-based businesses.”
“You must prioritise clarity and honesty over hype. No invented stats, no fake case studies.”
“You suggest structure, examples, and headlines, but always keep paragraphs short and in plain English.”
Copy that between projects and tweak it, rather than rewriting from scratch each time.
4. Nail your custom instructions
This is where most people quietly sabotage themselves. They leave custom instructions blank, or write something vague like “be helpful and concise” – then act surprised when the output sounds like generic corporate guff.
Fill each part in properly.
4.1 Who you are (role and expertise)
Tell ChatGPT who it’s working with so it can pitch things at the right level.
Example: “You’re helping a web designer and content strategist who’s worked with small businesses for 25 years. I know SEO basics and user experience, but I use AI as a thinking partner, not a truth machine.”
That one paragraph stops it over-explaining simple things and lets it give you more useful help where it matters.
4.2 Who you write for (audience)
If you don’t tell it who the reader is, it defaults to some imaginary “general audience”, which usually means bland. Spell out who you actually care about.
Example: “We write for small business owners and local service providers, often not very technical, usually juggling a lot. They’re smart, impatient, and allergic to jargon.”
Now it has a clear picture of who it’s trying to help, so examples, angles, and tone get sharper.
4.3 Your tone of voice
Without clear tone direction, you get American-style pep talk copy or stiff textbook paragraphs. Give it something like:
“Write in plain, conversational English. Short sentences, no buzzwords, no forced enthusiasm. Dry humour and a bit of bite are fine, as long as it’s honest and never cruel.”
This is also where you outlaw the phrases that make you want to throw your laptop.
4.4 What you hate
Being explicit about what you don’t want is one of the fastest ways to improve answers. Give it a blacklist.
Example: “Avoid corporate speak, invented statistics, fake urgency, or pretending we’ve done things we haven’t. Don’t talk about ‘leveraging synergies’ or any similar nonsense.”
You can be blunt here. The model will take you seriously.
4.5 What good output looks like
If you don’t define “good”, it’ll aim for something that ticks boxes (length, keywords, headings) instead of something you’d actually send to a client.
Tell it: “Good output is specific, structured, and usable. I should be able to paste 80% of it into a blog post or email without heavy rewriting. Use real-world examples and show your working when giving recommendations.”
You can even paste a piece of your own writing and say: “This is a strong example of my voice. Stay close to this.”
5. What you need to stop doing today
Here’s the uncomfortable bit. A few habits have to die.
5.1 Stop opening blank chats for repeat tasks
Every time you start from a blank chat, you erase context, tone, and previous corrections. You force the model back to factory settings – then blame it for not remembering you.
For anything you do regularly, work inside a dedicated Project. Treat a blank chat as a scratchpad, not home base.
5.2 Stop copy-pasting “magic prompts”
Those viral “just paste this” prompts are written to go viral, not to fit your business. At best they give you slightly less generic content. At worst, they ignore your values, push you towards spammy tactics, and can open you up to prompt-injection tricks – where hidden instructions in a shared prompt try to override your own.
You’re better off writing a simple, honest instruction in your own words and refining it over time.
5.3 Stop skipping thinking mode to save a minute
If you ask for a 1,500-word article in one go, you’ll usually get a 1,500-word mess. Break the work up and tell it to think.
For example:
“List 10 specific article ideas for [topic] that would genuinely help [audience]. Explain briefly why each one matters.”
“Pick the strongest idea and outline the article, including sub-headings and key points.”
“Draft the introduction and first section. Pause for review before continuing.”
Step-by-step prompting consistently produces tighter, more accurate content than “write me an article on X”.
5.4 Stop expecting AI to “know you”
ChatGPT isn’t psychic. It doesn’t remember your life, your brand, or last week’s meltdown about crap marketing jargon – unless you’ve given it persistent instructions and kept work inside the same Project.
If you’re reading its output and thinking “this sounds nothing like us”, that’s not the model being stupid. That’s the briefing being lazy.
6. A setup you can steal right now
Here’s a concrete template you can drop into a new Project today and tweak.
Project name: [Your Business] content and SEO assistant
Project description: You help me plan and write website copy, blog posts, and emails for our small business and local clients. You prioritise clarity, honesty, and long-term trust over clickbait or hype.
Copy and paste this into your Project instructions – then tweak for your own business:
“You’re working with a web designer and content strategist with 25+ years helping small, service-based businesses. I know SEO basics and use AI as a thinking partner, not a truth source.
We write for busy small business owners who are smart, impatient, and sceptical of marketing fluff. Use plain, conversational English. Short sentences. Dry humour is fine.
Avoid corporate speak, buzzwords, invented stats, fake urgency, and anything that sounds like an American sales webinar. Don’t fabricate experience, case studies, or data.
Good output is specific, structured, and practical. I should be able to use most of it as-is in a blog post, email, or web page. Show your reasoning and offer alternatives when you’re unsure.”
Upload two or three of your best pieces of writing and tell it: “These are strong examples of my tone and quality. Match them.”
From that point on, you’re not asking “ChatGPT” for help. You’re asking a configured assistant that already knows who you are, who you write for, and what bollocks to avoid.
The Questions You Were Going to Google Anyway
Do I need a paid ChatGPT account to use Projects?
Yes, Projects are available on ChatGPT Plus and above. The free tier gives you basic chat but no persistent workspace, custom instructions per project, or file uploads. If you’re using AI regularly for business, the Plus subscription is worth it.
How many Projects should I create?
One per distinct job type is the practical limit. If you find yourself creating a Project for every single client, you’re overcomplicating it. Think in job types: content, proposals, emails, SOPs. Four or five well-maintained Projects beats twenty neglected ones.
What files should I upload to a Project?
Your best existing writing, ideally two or three pieces that represent your actual voice. A strong blog post, a services page that gets enquiries, any brand or tone notes you already use. Don’t upload everything – upload the stuff you’d show a new writer on day one.
Will ChatGPT remember my instructions between chats?
Only if you work inside a Project. If you use a regular chat, it starts fresh every time with no memory of your tone, your clients, or your preferences. That’s the core problem Projects solve.
Is there a risk ChatGPT will make things up even with good instructions?
Yes. Better instructions reduce waffle and improve relevance, but they don’t eliminate hallucinations. Always check statistics, names, and specific claims before publishing. Treat it as a capable first draft, not a finished source of truth.
