Ingredients of AI

A great dish depends on more than the chef. It depends on the ingredients, the recipe, the freshness, and who's giving the instructions. AI is exactly the same: the final answer depends on the quality of the ingredients, not just the model.

Before we list those ingredients, let's lock in the trickiest ones — training, memory, and context — with an analogy you will never forget.

The Ghajini Analogy

If you've seen Ghajini, you remember Sanjay: after an accident he cannot form new long-term memories, so he survives using tattoos, photos, and notes. It turns out Sanjay is a perfect picture of how AI handles knowledge and memory.

The Ingredients of a Good AI Answer

Now the full recipe. Every useful AI answer is a combination of these building blocks. Change the quality of any ingredient and the answer changes with it.

Model — the brain

The trained engine that understands your request and generates the response (Layer 4 from the cake). Models differ in speed, intelligence, and cost — a small fast model is perfect for simple jobs; a large one earns its keep on hard reasoning. Picking the right one is a real skill.

Context — what it knows right now

Everything in the current conversation: your messages, the AI's replies, any files it has read. This is Ghajini's 15-minute memory — powerful but limited. When a conversation gets very long, the earliest parts start to fall out of the window.

Watch the window fill up below. The second tab shows the modern fix: when the limit is hit, today's AI summarises the older messages and keeps going, instead of simply forgetting.

Memory — what it remembers about you

Unlike context (which resets), memory persists across sessions — your preferences, your name, recurring facts. This is Sanjay's left-handed signature: a habit deep enough to survive the reset. It's why a well-set-up assistant can "remember" you prefer concise answers without being told each time.

Data — what it learned (and what it looks up)

Two kinds: the training data baked in (with its cutoff date), and live data it retrieves — web search for today's facts, or your own documents via RAG. Good answers depend on good, current data. This is exactly why you verify regulatory facts yourself: the training data can be stale or wrong.

Tools — what it can actually do

A model on its own can only produce text. Tools let it act: search the web, read a PDF, run code, query a database, send an email. This is the leap from a chatbot to Claude Code — an agent that uses tools to do real work on your machine.

Prompts — the instructions you give

The single ingredient you control most directly. A vague prompt produces a vague answer; a precise one produces a precise answer — same model, completely different result. Prompting well is the highest-leverage skill in this whole programme.

Reasoning — how it thinks through a problem

The model's ability to connect information and work through a problem step by step rather than blurting the first guess. Closely tied to temperature — the creativity dial. Low temperature = focused and predictable (what you want for code); high temperature = varied and creative (what you want for brainstorming).

Multimodality — text, images, audio, video, files

Modern AI isn't limited to text — it can read a scanned invoice, describe an image, transcribe audio. But here's the insight that matters most for you: code is not images or video — code is pure text. Where other tools spread effort across many data types, Claude concentrates on text, and code is the most demanding text of all. That is why this whole programme is built around Claude Code.

Guardrails — safety, privacy and rules

The controls that keep AI responsible — what it will and won't do, what data it protects, what policies it follows. For a CA this is not optional: never paste client PANs, financial details, or personal data into a public chat. Guardrails (and your own discipline) are what keep sensitive data safe.

Feedback Loop — how it improves

Good AI products learn from use — thumbs up/down, corrections, and continuous improvement. On your side, the feedback loop is the describe → run → verify → refine habit you've already been practising: each round teaches you to give better instructions.

The Recipe

AI Output = Model + Context + Memory + Data + Tools + Reasoning + Prompts (your instructions)

The model gets all the headlines, but the answer you actually receive is decided by every ingredient — and the one you control most is the last one. A world-class model with a vague prompt and no context gives a mediocre answer. A good model with sharp context, the right tools, and a precise prompt gives an excellent one.

Next: Going Deeper — tokens, context limits, temperature and hallucination in more detail, plus the difference between Claude.ai and Claude Code.