Be the Creator
You've been using other people's AI. Now build your own — a custom chatbot with your instructions and your knowledge, then agents that take action. Pick any concept. Press Esc anytime for this menu.
You don't build the brain
Here's the relief: you're not creating intelligence from scratch. The “brain” — a foundation model trained on the whole internet — already exists. Building your own AI means hiring that smart graduate and giving it a job, not raising a child from birth.
What you actually add
The foundation model
Companies spent millions and months training this. You get it for free or cheap — already smart, ready to direct.
The hiring analogy
The system prompt is the job description
The system prompt is a hidden set of instructions the AI reads before every chat — its role, rules and tone. Same model, totally different behaviour. Edit the instructions, ask the same question, and watch the bot change.
Try a preset job
Ask the bot
Same question, same model — the instructions decide the personality and rules. Edit the box and ask again.
Give it your knowledge
The model doesn't know your stuff — your prices, your rules, your notes. So you give it documents. When you ask a question, it finds the relevant snippets and answers from them. That's “retrieval” — handing the new hire your handbook.
The bot's knowledge base
Ask a question
Toggle a document off and re-ask — watch the bot lose the ability to answer it.
Grounded vs guessing
Ask about something specific to you. Without your documents, the model fills the gap with a confident guess — that's a hallucination. With them, it answers from real text and can point to the source. Flip the switch and see the difference.
Ask a question about “your business”
The lesson
The guess often sounds right — that's what makes hallucinations dangerous. Grounding in your docs is how custom AI stays trustworthy.
Same trick — meet AI
Personality & tone dials
The facts in an answer can stay identical while the voice changes completely. Part of building your AI is choosing how it sounds — formal, friendly, funny, brief. Pick a tone and watch the same answer get rewritten.
Tone
The question
“Is my order going to arrive late?”
Notice
The facts never change — your package is delayed one day. Only the delivery changes. That's the tone dial.
Test, find cracks, fix
No AI works perfectly first try. Building one is a loop: test it with real questions, find where it fails or goes off-script, tweak the instructions, and test again. Run the loop and watch the bot improve.
Run the build loop
Round 1 — first try
You test your bot. It works for easy questions but fails an edge case.
Same trick — meet AI
Where you actually build it
You don't need to code. The big assistants all have a no-code builder: you give it a name, instructions, and files — and you've got a custom AI. Click each to see what it's called and what's free.
Pick a platform
All of them give you
Examples teach it
Telling a bot the rules helps. Showing it examples helps more. Put a couple of “here's a good answer” samples in your instructions and the bot copies the pattern — format, length, style. Toggle the examples on and off.
Show examples in the instructions
Why it works
The model is a pattern-matcher. Good examples are the clearest instruction you can give — it imitates the shape you showed.
The opening message
A blank chat is intimidating. A good custom bot greets the user and offers a few “conversation starters” — buttons that show what it can do. It's the difference between a confused user and a confident one.
This bot is a…
A clear greeting + 3–4 starter buttons tells users exactly what to do. Click a starter to see it fill the chat.
Same trick — meet AI
What it should refuse
Part of building a good bot is telling it what not to do — stay on topic, don't give medical or legal advice, keep it family-friendly. Toggle the rules and watch how the same off-topic question gets handled.
Rules for your bookshop bot
Ask the bot
Turn rules on and re-ask the off-topic questions — the bot politely redirects instead of answering.
What it remembers
By default, a fresh chat is a blank slate — the bot forgets everything when you close it. “Memory” lets it carry facts about you between chats. Toggle memory and start a “new chat” to see the difference.
Memory
With memory off, the new chat has no idea who you are. With it on, it recalls — useful, but think about privacy.
Try to break it
Before you share your bot, attack it yourself. Users will try to make it ignore its rules, go off-topic, or say something embarrassing. Find the cracks first. Try each “attack” and see if a guarded bot holds.
Bot defenses
Try an attack
Guardrails off = it falls for tricks. On = it holds the line. This is how you harden a bot before launch.
Chatbot vs agent
A chatbot tells you what to do. An agent does it. Give both the same task — “book me a table for Friday” — and watch the difference: one writes instructions, the other takes the actions itself.
The difference
The chatbot hands you a to-do list. The agent uses tools — it searches, opens the booking site, fills the form, confirms. It acts in the world.
Same trick — meet AI
The agent loop
An agent works in a loop: think about what to do next, act by using a tool, observe the result — then repeat until the job is done. Step through a real task and watch the loop turn.
Task
“What's the weather where my 3pm meeting is, and should I bring an umbrella?”
Ready
Press “Next step” to watch the agent think, use a tool, see the result, and loop.
Tools an agent can use
An agent is only as capable as the tools you connect. Plug in a web search and it can look things up; a calendar and it can book; code and it can compute. Click a tool to see what it unlocks.
Connected tools
🔍 Web search
Now the agent can look up current facts — prices, news, opening hours — instead of guessing from memory.
More tools = more an agent can actually do. This is what “connectors” and “plugins” are.
When agents go wrong
An agent that takes actions can take the wrong ones — loop forever, misread a step, or do something costly. That's why real agents have guardrails and a human checking the risky moves. See what can break.
Failure mode
🔁 Stuck in a loop
The agent keeps retrying the same failing step forever, burning time and money. A step-limit guardrail stops it.
The fixes
✋ Human-in-the-loop for risky steps · 🚧 limits on steps & spend · 👀 logs you can review · 🔒 only the tools it truly needs.
Who's responsible?
You built it, so you're accountable for what it says and does — and for the data you feed it. Two things to get right before you share your bot with anyone.
Topic
Bottom line
Your custom AI is your responsibility. Don't upload secrets you don't control, check its answers, and be honest that it's AI.
Breaking a goal into steps
For a big task, a smart agent doesn't just dive in — it first writes a plan: a checklist of sub-tasks. Then it works through them one by one, ticking each off. Watch it turn one vague goal into a clear plan.
Goal
“Plan a birthday party for my friend.”
Why plan first?
Breaking a goal into steps means the agent can tackle each piece, check progress, and not get lost. It's the difference between “wander” and “execute.”
Same trick — meet AI
Connecting to the world
How does an agent actually reach a calendar or a shop? Through connectors — standard “plugs” (APIs, and newer standards like MCP) that let your bot talk to outside services safely. Click a service to see it plug in.
Plug in a service
📅 Google Calendar
A calendar connector lets your bot read your schedule and create events — with the permissions you grant.
You decide what each connector is allowed to do. A connector is a door — you choose how wide it opens.
A team of agents
Big jobs go better with a team. Instead of one agent doing everything, you can have specialists — a researcher, a writer, a checker — each great at one thing, passing work down the line. Run the team on a task.
Task
“Write a short, fact-checked post about honeybees.”
The team
Smart vs fast vs cheap
Not every job needs the biggest brain. Bigger models are smarter but slower and pricier; smaller ones are fast and cheap. Part of building well is picking the right model for the task. Slide between them.
Model size: balanced
Balanced model
A good default — smart enough for most tasks, reasonable speed and cost.
Rule of thumb
Simple, high-volume tasks → small/fast/cheap. Hard reasoning → big/smart. Many builders mix: a small model for easy steps, a big one only when needed.
Publish & share
Once your bot works, you can publish it — share a link, list it in a store, or let others remix it. This is where “building” turns into “reaching people” (and, on Day 5, earning). See the journey.
Step
Build & test
You create the bot, give it instructions and knowledge, and test it privately until it's good.
Looking ahead
A useful bot others rely on can become a product. Day 5 covers turning this into income.