Building Custom GPTs
How to create your own specialised AI assistants — and three of mine you can try.
One of the most powerful and underrated features in ChatGPT is the ability to build your own custom GPTs. These are specialised AI assistants you design yourself, trained on your own knowledge, speaking in whatever voice you choose, and focused on exactly the task you need.
Think of it this way: ChatGPT out of the box is a generalist. A custom GPT is a specialist you have hand-built. You decide what it knows, how it talks, and what it does.
I have built several custom GPTs and showed them to a friend who found them genuinely fascinating. So I thought I would share three of my favourites here, explain how I built each one, and then walk you through building your own from scratch.
Three GPTs I Built (and Why They Work)
If you are Australian (or just enjoy watching a politician with a gift for language tear strips off someone), this one is for you. Former Prime Minister Paul Keating was famous for his withering put-downs — calling the Australian Senate "unrepresentative swill," telling an opponent he would "do you slowly," and dismissing a rival's economic credentials with the kind of precision that left permanent marks.
This GPT searches the internet for today's most talked-about Donald Trump news, then delivers commentary in Keating's unmistakable voice. The results are frequently hilarious, occasionally savage, and always articulate. Keating's style — that mix of deep policy understanding, cultural sophistication, and absolute contempt for mediocrity — turns out to be the perfect lens for commentary on modern American politics. He would have had a field day with this era, and now, in a sense, he can.
This one started as a joke at my book club — Tough Guy Book Club (TGBC) — and turned into something people genuinely love using. It is, at its core, a cheerfully grim mortality calculator crossed with a book recommendation engine. And it does not pull punches.
Here is how it works. It asks your age and gender, then uses Australian average life expectancy data to calculate how many years you have left. Then it asks how many books you read per year. The maths is simple and brutal: years remaining multiplied by your reading pace equals the total number of books you will ever read. It gives you two numbers — "low effort you" (your current pace) and "best version of you" (one book a week). The gap between those two numbers is designed to sting a little.
Then it builds your reading list for the coming year. There are non-negotiable rules. One Hemingway novel is mandatory every year — you will complain about this, everyone does, but at TGBC we have read most of them and it is tradition. Three books about death and ageing are also compulsory, because if you are going to face the clock you might as well read about it. The remaining slots are filled based on genres you actually enjoy, suggested one at a time so you can approve or reject each pick. If you have already read something, it swaps it out. It avoids obvious bestseller picks unless you specifically ask for them, and it can go either "mainstream but overlooked" or "weird, obscure, slightly unhinged" depending on your taste.
Once your list is locked in, it checks Australian booksellers — Booko, Dymocks, Booktopia — for prices, stock, and direct links. No guessing, no dodgy information.
And then it roasts you. Based on your age: if you are young, it tells you to get moving. If you are old, the clock is ticking. If you are very old, it asks why you are still on your phone. The whole vibe is dry, a bit dark, zero motivational fluff — like a mate at the pub telling you to stop wasting time.
This is the practical one — and honestly, the one I use most often in real life. It is not a general-purpose assistant. It is a dedicated negotiation strategist — a personal coach that works for any situation: buying a car, asking for a raise, resolving a dispute with a neighbour, renegotiating a contract, pricing a client engagement, handling a supplier who just jacked up their rates, or trying to talk your way out of something you probably should not have done in the first place.
You describe your situation and it gets to work. It follows a structured method: first a summary of the issue, then the key negotiation principles that apply, then structured guidance, practical examples with actual language you can use, and finally your next steps. It does not give you vague advice like "be confident" — it diagnoses your negotiation the way a strategist would. It identifies your BATNA (best alternative), your reservation point, the other side's likely interests and fears, any power imbalance, timing pressure, relationship importance, and possible trade-offs. Then it builds you a concrete game plan: opening moves, concession sequences, specific phrases to say out loud, and rebuttals for when the other side pushes back.
The knowledge base is deep. It covers Harvard's principled negotiation method (BATNA, ZOPA, interest-based bargaining), Chris Voss's FBI hostage negotiation techniques (mirroring, labelling, tactical empathy), anchoring strategies, coalition building for multi-party deals, cross-cultural negotiation and cultural intelligence, emotional intelligence tactics, power-based strategies, how to handle time pressure and artificial deadlines, information asymmetry, trust-building, and concession planning. Beyond the frameworks, it has domain-specific playbooks for salary negotiation, real estate, vendor and procurement deals, client pricing, contract renegotiation, co-founder equity disputes, and the subtle art of domestic negotiation when you have clearly done something wrong.
It also asks clarifying questions when your situation is incomplete, because bad advice from half the story is worse than no advice at all. And it can work at whatever level you need — high-level strategy, exact scripts and wording, role-playing the other side, stress-testing your BATNA, or turning your plan into a negotiation checklist. You can even use it after a negotiation to do a post-mortem: what happened, where leverage shifted, which tactics were used against you, and what to do differently next time.
How to Build Your Own Custom GPT
You do not need to be a programmer. You do not even need to be particularly technical. If you can write a clear brief and have an idea for something useful, you can build a custom GPT in an afternoon. Here is exactly how.
Start with a clear idea
Before you open ChatGPT, write down in one or two sentences what your GPT should do and who it is for. "A GPT that helps me draft professional emails in my company's tone" is a good start. "A GPT that does stuff" is not.
The best custom GPTs solve a specific, recurring problem. Think about tasks you do repeatedly where you always end up explaining the same context to ChatGPT.
Open the GPT Builder
In ChatGPT, click your profile picture, then My GPTs, then Create a GPT. You will see two tabs: Create (a conversational wizard) and Configure (manual setup). I recommend going straight to Configure — it gives you more control and you will understand what you are building.
Write your system instructions
This is the most important step. The instructions tell the GPT who it is, how it should behave, and what it should do. Write them as if you are briefing a very capable new employee.
Be specific about tone ("speak in a direct, professional manner"), behaviour ("always ask clarifying questions before giving advice"), and boundaries ("never provide medical or legal advice — suggest consulting a professional instead").
For my Keating GPT, the instructions run to several pages covering his rhetorical style, political positions, common phrases, and how to structure a response. For simpler GPTs, a few paragraphs will do.
Build your knowledge base
This is where custom GPTs become genuinely powerful. You can upload files — PDFs, text documents, spreadsheets — that the GPT will search when answering questions. This is called RAG (Retrieval-Augmented Generation), and it means your GPT can draw on specific knowledge that base ChatGPT does not have.
For the Keating GPT, I uploaded transcripts and quote collections. For the Negotiation Advisor, I uploaded the tactics database I built with Manus and Perplexity. You could upload your company's style guide, your product documentation, your recipe collection — anything that makes the GPT smarter about your specific domain.
Configure capabilities
Choose what your GPT can do beyond conversation. You can enable Web Browsing (so it can search the internet — essential for my Keating GPT's daily news commentary), DALL-E Image Generation, and Code Interpreter (for data analysis and calculations — used in the Tough Guy Book Club mortality calculator).
Only enable what you actually need. Every capability you turn on is a potential distraction from the GPT's core purpose.
Test relentlessly
Use the preview panel to test your GPT with real scenarios. Try to break it. Ask it questions outside its scope. Give it edge cases. The preview panel shows you exactly what the GPT will say, and you can go back and refine your instructions based on what does not work.
I went through dozens of iterations on the Keating GPT alone — refining the voice, adjusting how it selects news stories, and tuning how it balances humour with substance.
Publish and share
When you are happy, you can publish your GPT as private (just for you), shared via link (anyone with the link can use it), or public (listed in the GPT Store for anyone to discover). You can change this at any time.
Advanced Techniques
Once you have the basics down, there are a few techniques that separate a decent custom GPT from a genuinely impressive one.
Personality is everything. The most engaging GPTs have a distinct voice. Do not settle for generic AI assistant speak. Write sample responses in your instructions showing the GPT exactly how you want it to sound. Include phrases it should use, attitudes it should express, and styles it should avoid.
Use Manus or Perplexity to build your knowledge base. One of the hardest parts of building a GPT is assembling the knowledge it needs. I used Manus to autonomously research negotiation tactics across dozens of sources, and Perplexity to verify and deepen specific areas. These tools can compile research in hours that would take you days manually.
Structure your uploaded documents well. The GPT searches your uploaded files using RAG, so structure matters. Use clear headings, consistent formatting, and logical organisation. A well-structured document produces dramatically better results than a messy one.
Set boundaries explicitly. Tell the GPT what it should not do, not just what it should do. "If the user asks about [topic outside scope], politely redirect them to [appropriate resource]" prevents your GPT from wandering into territory where it will give poor answers.
Iterate based on real use. The first version of your GPT will not be perfect. Use it yourself for a week, note where it falls short, and refine the instructions. The best GPTs are the ones that have been polished through actual use.
Ideas to Get You Started
Not sure what to build? Here are some custom GPTs that real people have found genuinely useful:
Personal writing editor — upload your best writing samples and have a GPT that edits new work to match your established style and voice.
Meeting preparation assistant — upload your company's strategic priorities and have a GPT that helps you prepare talking points, anticipate questions, and draft agendas for any meeting.
Recipe curator — upload your family's recipe collection and dietary requirements. Ask it to plan weekly meals, suggest variations, and generate shopping lists.
Study companion — upload your course materials and textbook notes. Have a GPT that quizzes you, explains concepts in different ways, and tracks what you have covered.
Travel planner with your preferences baked in — upload your past trip itineraries and preference notes. Get recommendations that already know you prefer walkable cities, hate guided tours, and always want a local breakfast spot.
The possibilities are genuinely endless. The constraint is not the technology — it is your imagination and willingness to iterate.