🎤 Interview Cheat Sheet
Read this on the train before the interview. Every answer is written to be said out loud — two to four sentences, no jargon you can't back up.
How to use this
Don't memorise it word for word — read each answer twice, then say it in your own voice. Interviewers aren't testing whether you can define a transformer; they're checking that you've actually used this stuff, that you're honest about its limits, and that you can explain it to a colleague without the buzzwords. Concrete beats clever, every time.
The questions
“So — what actually is AI?”
AI is software that does things we'd normally say need human intelligence — understanding language, recognising images, making decisions. The AI in the news today has mostly learned from huge amounts of data rather than being programmed rule by rule. I think of it as a very fast, very capable junior colleague: brilliant at drafts and analysis, but it still needs someone checking its work.
“What's an LLM? How does ChatGPT actually work?”
A large language model is a neural network trained on enormous amounts of text to do one deceptively simple thing: predict the next word-chunk, called a token. Do that billions of times with billions of internal parameters and it turns into answering questions, writing copy and holding conversations. ChatGPT and Claude are the friendly chat apps — the LLM is the engine inside.
“What's a token, and what's a context window?”
A token is the chunk of text a model reads and writes — roughly three-quarters of an English word. The context window is the model's working memory: how much it can consider at once, and if a long conversation overflows it, the oldest bits fall out. It matters commercially too, because API usage is priced per token.
“What's the difference between ChatGPT, Claude, Codex and Claude Code?”
ChatGPT is OpenAI's chatbot and Claude is Anthropic's — both are chat apps wrapped around large language models. Codex and Claude Code are those same companies' coding agents: instead of just chatting, they sit in your terminal or editor, read your files, write code and run commands. So the split is chat apps for conversation, agents for actually doing the work.
“What's an AI agent?”
An agent is AI in a loop with tools: you give it a goal, and it plans, acts — runs a command, calls an API, edits a file — checks the result, and decides its next step until the job's done. It's the difference between asking a chatbot “how do I build a website?” and telling an agent “build me a website” and watching it happen. I've done exactly that on a course I took.
“What's MCP? I keep seeing it mentioned.”
MCP is the Model Context Protocol — an open standard for plugging AI models into outside tools and data, which is why people call it the USB-C of AI. An MCP server wraps one service, say Google Analytics or a database, in a standard shape that any capable AI can then use. Before MCP every integration was custom; now they snap together.
“Where could AI help in a marketing or PR team?”
The obvious wins are first drafts — press releases, social posts, campaign variants — plus summarising coverage and turning analytics into plain-English reports. The more interesting step is agents: connect one to Google Analytics or an ads account and it can pull the numbers, spot the story and draft the report itself. The human stays on strategy, relationships and final sign-off.
“What are AI's limits? What's a hallucination?”
The big one is hallucination: a model will state something false with total confidence — invented statistics, citations, quotes — because it's trained to produce plausible text, not verified truth. Models also have knowledge cut-offs and know nothing about your business unless you give them context. So: a brilliant drafting and analysis tool, never an unsupervised source of facts.
“How do you verify AI output?”
My rule of thumb: the more specific and checkable a claim, the more I verify it — names, numbers, dates, quotes and links all get checked against the source. For anything customer-facing I keep a human review step, and I ask the model to say where things came from, which makes checking much faster. Treat it like work from a bright new intern: usually excellent, always reviewed.
“What would you automate first — and what guardrails would you set?”
I'd start with something high-volume, low-risk and easy to check — say a weekly analytics summary or first-pass media monitoring — so the win comes quickly and mistakes are cheap. For guardrails: a human approves anything that gets sent or published, the AI can't spend money or delete things without sign-off, and we review its output regularly rather than assuming it's fine. Prove value on the boring stuff, then move up.
“Open versus closed models — what's the difference?”
Open-weights models like Llama or Kimi publish their trained parameters, so you can download and run them on your own hardware — great for control, privacy and cost. Closed models like GPT-5 and Claude are reached through the maker's API, and usually offer the strongest capability with zero infrastructure to manage. Most teams start closed for quality and simplicity, and look at open when privacy or scale demands it.
“What's changed in AI recently?”
The headline shift is from chatbots to agents — AI that doesn't just answer questions but takes a goal and works through it with tools. Standards like MCP mean models now plug into real business systems, and open-weights models keep closing the gap on the closed ones. In short: less novelty, more actual work getting done.
“What's your favourite prompting technique?”
Giving the model a role and a format: “you're a marketing analyst — summarise this in five bullet points for a client” beats “summarise this” every time. I'll also show an example of what good looks like, because models are superb mimics. And for anything big, I put the context in first — a model can only work with what it can see.
“Tell me about something you've built with AI.”
I built and shipped a real website using Claude Code — described what I wanted in plain English, reviewed what the agent produced, and iterated until it was live. I also set up a personal AI agent I can message from Telegram, and connected AI to analytics data so it could pull the numbers and explain them in plain English. Small projects, but end-to-end: brief, build, verify, ship.
“Is AGI just around the corner?”
AGI means AI that matches humans across most intellectual work — and honestly, there's no agreed definition or test, and expert predictions range from a few years to decades. I focus on what today's systems demonstrably do, which is already transformative in a workplace, without over-claiming. Powerful tools, real limits, moving fast — that's the honest answer.
Power vocabulary
Ten words that make you sound fluent — one line each, drop them naturally or not at all.
Do / don't
| ✅ Do | ❌ Don't |
|---|---|
| Give concrete examples of things you've actually done | Serve buzzword soup (“we'll leverage synergistic AI at scale”) |
| Be honest about limits — hallucinations, verification, human review | Claim AI is infallible, or that AGI has arrived |
| Name the human in the loop for anything published or sent | Suggest automating customer-facing output unsupervised |
| Distinguish the app from the model (ChatGPT is the app; GPT-5 is the model) | Use “AI”, “algorithm” and “model” interchangeably |
| Say “I'd want to verify that” when you're not sure | Bluff about a tool you've never actually used |
One last thing ✈️
You know more than you think you do. Slow down, give one concrete example per answer, and let them ask the follow-up. Good luck — you've got this.