Everything you need from Mat's session, turned into something you'll actually use. The secret sauce, the tools, the prompts, the traps to dodge. One page. Your cheat code.
If you take one thing away from the session, take this. It's cheesy, it's true, and it's the single biggest predictor of whether you'll get good results from AI.
The best AI users aren't the most technical people in the room. They're the most curious. The ones who keep asking "can you help me with this?" until something useful falls out.
The old world: bash your head against a spreadsheet for an hour, watch three YouTube videos, try a formula, break it. The new world: the tool talks back. Ask it how to do the thing. Ask it to write the formula for you. Ask it why its last answer was wrong.
If you're good at what you do, you'll get better results than someone who isn't — because you'll know the right questions to ask and you'll spot when it's fibbing.
Strip away the hype and the buzzwords and this is what you're really looking at. Four things to know, then you'll understand why it behaves how it does.
Large Language Models (LLMs) — your ChatGPT, Claude, Copilot, Gemini — sit in big data centres and have consumed almost all publicly available text.
It spots patterns in how words combine across languages and contexts. That's why it can write across domains, and why it's often strong on translation.
Given what's come before, it guesses what likely comes next — word by word, paragraph by paragraph. That's the engine.
Reinforcement learning with human feedback (RLHF) — thousands of people scored responses as good or bad. That's the leap that made it genuinely useful.
Because it isn't calculating anything. When you ask it for 500 × 72 − 45 × 100, it's predicting what the next characters look like based on similar maths it's seen before. Use a spreadsheet for arithmetic. Use AI for pattern-spotting, categorising, drafting, and explaining. Different jobs.
Every platform calls them something different. ChatGPT calls them GPTs. Gemini calls them Gems. Copilot calls them Agents. Claude calls them Projects. Same idea in all cases.
One chat, one topic. Don't ask your marketing chat for a beef stew recipe. Context bleeds, answers get worse. Keep conversations specialised and you'll get consistently better results — same as you would with a human you'd hired for a specific role.
AI can't read your mind. Yet. Until it can, feed it these four ingredients in most prompts and watch your results jump.
Who should it be? "Act as an experienced office manager…" sets the vocabulary and depth.
Be specific. Not "help with my email" — "rewrite this under 150 words, keep the ask in line one."
Who you are, who reads this, what it's for, any constraints. AI doesn't know your world unless you tell it.
How do you want it? Table? Bullet list? Three options? 200 words max? Examples help more than you'd think.
RLHF leans the model toward safe, familiar responses. These three moves pull it out of vanilla mode.
Covered in the session, in order of how often they came up. Start with one or two — don't try to learn all of them at once.
Consistently excellent writing and reasoning. Handles long documents and templates beautifully. Claude's design and artifact features can generate whole websites and visual assets from a prompt.
Built to search the live web before answering. Lets you pick which underlying model to use (Sonar, GPT, Claude, etc.) and has an academic mode for research papers.
Agent Mode (paid tier) opens its own browser and takes action — great for pulling structured research into a spreadsheet. Significantly fewer hallucinations than a normal chat because it works from live sources.
Lives in Word, Excel, Outlook, Teams. Data stays in your Microsoft tenancy. Now lets you pick underlying models (including Claude). Build Agents scoped to a specific SharePoint folder — don't let them search everything.
Upload big documents (PDFs, papers, transcripts). Grounded to your sources — much lower hallucination rate. Generates infographics, quizzes, and audio overviews. Brilliant for turning old lecture handouts into living study material.
A browser that takes control on your behalf. Tell it "research construction firms in the South East and put them in a sheet" and it does it. Powerful, but keep it away from anything sensitive.
Has a mode that spins up multiple specialised agents — each does its piece, then they combine. A glimpse of where orchestrated AI is heading: lots of narrow specialists beats one generalist.
British language model with UK data residency. Relevant if your contracts — public sector, critical infrastructure — forbid US data centres. Keeps your data on-shore.
Call AI directly from a cell. Classify sentiment on 500 reviews, generate summaries, enrich a company list with research — then drag the formula down. Hours turn into seconds.
Ask once, have it run every Monday at 9am. News briefings, inbox summaries, repeated research. Three dots under any reply in Copilot, or just tell ChatGPT "do this every weekday at 9am."
Run on your phone or laptop. No data leaves your device, costs nothing, privacy by default. Mat estimates 60–80% of everyday queries could be handled by one of these.
Pair with a site built by Claude or similar — drag-drop the file, you've got a live website for free. This is what "AI killing SaaS" looks like in practice: cheap, fast, yours.
Knowing these stops you falling into the traps everyone else falls into. None of them are reasons not to use AI — they're reasons to use it properly.
Because it's predicting the next word, if it gets one bit wrong it'll just keep going — and then defend the wrong answer if you push back, before flipping to a different wrong answer. The fix: ground it in source material. Upload the PDF. Point it at a specific webpage. This is called Retrieval Augmented Generation (RAG) and it dramatically cuts hallucinations.
On free or standard paid ChatGPT, Claude, Gemini — your conversations are used to improve the model unless you go into settings and turn it off. Enterprise tiers and M365 Copilot are generally opt-out by default. Never paste anything sensitive into a consumer chat. Shared conversation links can become publicly indexed on Google.
When you use an agent that browses the web, it reads the HTML. A malicious page can hide instructions in white-on-white text that tell your agent to do something else entirely. Sam Altman has admitted this may never be fully solvable. The rule: don't give agentic tools access to anything you'd mind losing — no banking apps, no sensitive systems.
Someone convinced their customer-facing chatbot that the deal was legally binding. The internet had a field day and the bot came down fast. Moral: test chatbots internally for weeks before you put them in front of customers, and never give a public bot access to sensitive data or decisions it could be talked into making.
Once a conversation gets big — roughly past 40% of the model's context window — quality drops. Models remember the start and the end better than the middle, same as humans. When a chat you've loved starts giving worse answers, that's your cue to either start fresh or upgrade it to a proper Agent/GPT/Project with a fixed set of instructions.
Most models are trained to be agreeable. Say something wrong and they'll often nod along. Tell it up front: "push back on me, disagree if I'm wrong, don't just validate" — you'll get more honest answers. Useful especially for strategy work where you're looking for holes, not hugs.
Tick them off as you go — progress saves in your browser. Small steps, in the right order. This is the sequence that builds a real AI habit.