Build Something Real · AI workshop

Step 0 · do this before you arrive

Five minutes, a few days early

5 minPre-work

Everything on the day happens inside the app you already have. Nothing to install, nothing to download. But three small things are much easier to sort out now than in a room with forty other people.

By the end of this page

  • You can sign in on the machine you're bringing
  • You can see the company connection in your settings
  • You've sent us a screenshot if either of those didn't work

1. Sign in on the laptop you're actually bringing

Not the one on your desk — the one that will be in the room. Use your work account, not a personal one. The company connection only exists on the work account, and a personal account will look almost identical while quietly doing less.

2. Check the company connection is there

Open your settings and look for the list of connected apps. You're looking for one labelled [COMPANY CONNECTION NAME]. You don't need to do anything with it yet — just confirm you can see it.

You should seeIt in the list, with a label marking it as a custom company app. If it's switched off, that's fine — leave it. We turn it on together.
I can't see it at all

Most likely one of three things, in order of likelihood:

  1. You're signed into a personal account rather than your work one. Check the account name.
  2. Your access hasn't been granted yet — this is common and takes minutes to fix, but not on the morning of.
  3. You're looking in the wrong place. The menus have been renamed twice this year.

Send us a screenshot of your settings screen. Don't spend more than five minutes on it — that's what we're here for, and finding out now costs nothing.

3. That's it

Genuinely. Don't prepare anything else, don't read ahead, and don't arrive with a project you've already half-built — you'll get more out of the day starting fresh.

If you already have an idea for something you'd like to build, bring it. If you don't, that's completely normal and the day is built for it — you'll find one in step 5 without having to think about it beforehand.

Step 1

Put something on the internet in the next fifteen minutes

20 minEveryone ships

We're going to do the scariest part first, while it's still trivial.

Most people expect a day like this to end with a nerve-wracking "and now we publish it" moment. We're inverting that. In a few minutes you'll have a real web address, live, that you could send to a colleague — with almost nothing on it. After that, publishing is just a thing you've already done.

By the end of this step

  • You have a live web address that belongs to you
  • You've seen how little it takes
  • You've sent it to the person sitting next to you

Say what you want

Start a new chat. Paste this in, changing only the words in the middle:

Do this

Make me a one-page website with a big heading that says "Hello, I'm [YOUR NAME]" and one sentence underneath saying what I do at work. Keep it simple. Then publish it to [INTERNAL HOST] so I have a live link.

That's the whole instruction. You don't need to know what it's building or how it gets there.

You should seeA link you can click, ending in our internal address. Open it. That page is yours and it is genuinely live — anyone here can visit it right now.

Checkpoint

Blue sticky on your laptop lid if your page is live. Yellow if it isn't. If you're on yellow, someone will come to you — you don't have to say anything or put your hand up.

Now send it to someone

Turn to the person next to you and send them the link. Open theirs. This takes thirty seconds and it is not filler — it's the first moment the thing stops being hypothetical.

It made the page but didn't publish it

Common. It built the thing and stopped short. Just ask again, in a separate message:

Now publish that to [INTERNAL HOST] and give me the link.

This is worth noticing rather than skipping past: it did most of what you asked and told you it was done. That happens, and step 2 is largely about catching it.

It asked me something I don't understand

You're allowed to say so. Try: "I don't know, you pick something sensible." It will. You are not required to have opinions about anything technical today, now or later.

Worth saying out loud: the model is inconsistent. It will confidently produce something wrong several times today. When that happens it is a property of the tool, not a comment on you — and spotting it is the actual skill we're here to build.

Step 2

Change it by describing what's wrong

25 minThe core loop

You now own something you can look at. That changes what you're able to do — because you're a much better critic of a thing that exists than of a thing you're imagining. An idea in your head always looks finished. On screen, you can see what's actually wrong with it.

The whole rest of the day is one loop, repeated:

The loop

  • Describe — say what you want
  • Look — actually open it and look
  • Change one thing — name a single change, not five
  • Publish — put it live again

One change per message

This is the single most useful rule of the day, and it's counterintuitive. When you ask for several things at once, you'll often get some of them — and be told you got all of them. Asking for one thing at a time is slower per message and far faster overall.

Do this — three times, separately

Make three changes to your page, one message each, looking at the result between each one. For example:

Change the background to something warmer.
Add a short list of three things I'm working on this quarter.
Make the heading smaller and the list bigger.

Publish after each one and open the live link. Yes, really open it each time.

You should seeThe live page changing each time. If the live page looks stale, you probably haven't republished — the preview and the live site are not the same thing, and this catches almost everyone once.

Now break it on purpose

This matters more than it sounds. You need to have gone wrong and come back before it happens for real.

Do this

Redesign the whole page completely differently. Go wild.

Look at it. You'll probably hate it. Now get back:

That's worse. Go back to how it was before that change.

Checkpoint

Everyone should have gone wrong and recovered at least once. Blue if you've broken it and got back. Yellow if you're stuck in a bad version.

The two-tries rule — worth adopting today

If you've asked for the same change twice and it still isn't right, stop asking again. Go back to the last good version and describe the outcome differently, or start that piece over.

People get stuck in loops of re-asking, each attempt a bit further from what they wanted. Two tries, then change tack. It feels like giving up; it's the opposite.

If you'd rather talk than type

You can dictate instead of typing, and some people find describing a change out loud much easier than writing it. Others find talking to a laptop in a room full of colleagues deeply strange. Both are fine — nothing today depends on it, so use it or ignore it entirely.

Step 3 · the important one

Point it at your actual work

30 minCentrepiece

Up to now you've been working with a blank page. That's the version of this that most people have tried, and it's why most people conclude it's a clever toy.

This step is the difference. We connect the app to the systems your work already lives in, so it stops guessing and starts answering from your real context.

By the end of this step

  • The company connection is switched on for you
  • You've asked it something about your own real work
  • You know what it can and can't see

What this actually is

There's a shared standard for connecting a chat app to systems it wasn't trained on — think of it as a common plug. Someone here has already built the company's plug and wired it to our tools. All you do is switch it on.

It doesn't make the model smarter. It gives it a supervised way to look things up and make changes on your behalf, using your access and your permissions.

The question everyone asksIt can only reach what you could already open yourself. It doesn't get a copy of your chat history, and it can't see anything your own account can't. If you can't open a document today, neither can it.

Switch it on

Do this

Go to your settings, find [COMPANY CONNECTION NAME] in the connected apps list, and turn it on. You'll be asked to sign in and approve it — that's the normal one-time permission step.

Then start a fresh chat and select the connection for your message.

Two things that will confuse you, so here they are in advance. First, choosing the connection applies to the message you're sending, not the whole conversation — you may need to pick it again later. Second, we'll start a new chat for each exercise today. Each new chat asks for approval again. That's expected, not a fault.

Ask it something only you would know

Do this

Pick whichever of these is real for you:

What are the tickets currently assigned to me? Summarise what state they're in.
Find the most recent spec I was involved in and tell me what's still unresolved in it.
Catch me up on what's happened this week on the things I'm involved in.
You should seeAnswers about your actual work — real names, real tickets, real threads. This is usually the moment the room goes quiet.

Checkpoint

Blue if you got an answer about your real work. Yellow if not. This is the one step where nobody moves on without it working — if you're on yellow, we come to you before we continue.

It answered, but the answer is wrong

Good catch — genuinely. This is the most valuable thing you'll notice today.

It can retrieve the wrong thing and describe it with total confidence. Ask it which specific items it used, and go and look at one of them yourself. Verify in the actual system, not in the chat. That habit is worth more than any prompt you'll learn today.

Nothing happened, or it says it can't access anything

Try in order: check the connection is actually toggled on; start a completely fresh chat and select it again; confirm you're on your work account.

If it's still nothing after two minutes, put up a yellow sticky. Don't keep re-asking — that never fixes this particular problem.

Worth knowing: it reads whatever it's given

The model reads ticket text, comments and messages — including things written by people outside your team, or outside the company. Anyone who can file a ticket can put words in front of it.

So treat anything it does on your behalf the way you'd treat a new joiner acting on your instructions: useful and fast, and you still check the important ones before they go out.

Step 4

Turn a pile of mess into something you'd actually send

30 minPairs

This is the most transferable thing today, and it's the one you'll use on Monday whether or not you ever build another website.

You'll practise on a mess we've supplied — contradictory notes, three versions of the same document, a thread that never resolved. Then you'll do it on your own.

By the end of this step

  • You've turned a genuine mess into a one-page brief
  • You've made it disagree with you and defend its answer
  • You've done it once on your own material
Pair up for this one. Everyone pairs, not just people who want to — one person drives, the other watches and questions, and you swap halfway. Do not pair with someone who reports to you, or who you report to.

First, on the supplied mess

Do this

Open [SAMPLE PACK LINK] — a small pile of realistic material for a project that has gone sideways. Attach the lot, then:

These are all the notes and documents I have about this project. They contradict each other in places. Give me a one-page brief: what we're actually trying to do, what's been decided, what's still open, and what's contradictory. Flag the contradictions specifically rather than smoothing over them.
You should seeA brief that is roughly right and confidently wrong in one or two places. Both halves matter — find the wrong parts with your partner.

Now push back on it

Most people accept the first answer. The value is in the second and third.

Do this — one at a time

You said something was decided that I don't think was ever agreed. Which document did that come from? Quote it.
Rewrite this for someone who has ten seconds and only needs to know what to do next.

Checkpoint · swap drivers

Whoever hasn't been typing, take the keyboard now. Blue when you've swapped and got a brief you'd be willing to send.

Now on something of yours

Do this

Using the company connection, point it at something real and unresolved that you own:

Look at [YOUR THING — a project, an epic, a spec, a channel] and give me the same one-page brief: what we're doing, what's decided, what's open, what's contradictory.
Stretch — make it a document your team can use

Ask it to write the brief back into the actual system as a comment or a page. Then go and look at it there.

Check what it actually wrote before you tell anyone. It will occasionally report success on something it didn't do.

Step 5

Let it interview you about what to build

20 minChoose your project

Now you pick what you're building for the rest of the day. If you already know, skip straight to the brief. If you don't — which is completely normal — you're two prompts away from knowing.

You are not expected to have arrived with an idea. A ready side project is not a normal state for anyone with a real job. What follows finds one in your existing work rather than asking you to invent one.

Find the thing worth building

Do this

Interview me about my working week. Ask me one question at a time — no more than six questions total. You're looking for something repetitive, annoying, or slow that a simple internal web page could fix. At the end, suggest three specific things I could build today in about an hour, and say which one you'd pick and why.

Answer honestly rather than impressively. It's looking for tedium, and tedium is where the value is.

You should seeThree concrete suggestions, at least one of which makes you think "oh, that would actually be useful."
Nothing it suggested appeals — pick from these instead

Perfectly reasonable. Any of these is a real thing people here have wanted, and none is a consolation prize:

  • A status page for your area that answers the questions people keep asking you in DMs
  • A one-page explainer of how your team works, for new joiners and cross-team partners
  • A decision log — what was decided, when, and why, so it stops getting relitigated
  • An onboarding checklist for a process only you currently know
  • A live summary of your current work, pulled from the systems, that you can send instead of writing an update

Get the plan out of your head

Before it builds anything, make it tell you what it thinks you asked for. Cheapest thirty seconds of the day.

Do this

Before you build anything: describe back to me, in plain language, exactly what you're going to make. What's on the page, who it's for, and what it does. Don't write any of it yet.

Checkpoint

Blue when you have a plan you agree with. Not a good plan — one you agree with. You'll change it anyway.

Scope warning, and it's a real number: people reliably build about a quarter of what they expect to in a fixed session. Take whatever you're picturing and cut it to the smallest version that would still be worth sending to someone. You'll have time to add to it. You won't have time to rescue it.

Step 6

Build the thing

60 minPairsHelpers circulating

This is the long block. Same loop as step 2, bigger thing: describe → look → change one thing → publish.

Where you're heading

  • Something live at a real address, early and ugly
  • Changed at least five times after you first saw it
  • Good enough that you'd send it to one specific person

Get it live before it's good

Do this first

Build the simplest possible version of that plan and publish it to [INTERNAL HOST] straight away. Don't make it nice yet — I want to see it live first.
You should seeSomething you're slightly embarrassed by, at a real web address. That's exactly right. Everything from here is improvement, and improvement is easy.

Then work in small changes

  • One change per message. Look at the live page between each.
  • Two failed attempts at the same thing → go back and take a different run at it.
  • Swap driver every fifteen minutes. Whoever is quieter takes the keyboard first.
  • If your partner is stuck, resist typing on their laptop. Ask before you touch it.

Checkpoint · halfway

Blue if it's live and you've changed it at least twice. Yellow if you're not live yet. Being on yellow at halfway is normal and fixable — it's only a problem if nobody knows.

Stretch — make it pull from real systems

If your page would be more useful with live information in it, connect it to the real thing rather than typing the content in by hand.

Instead of me writing this list by hand, pull it from [the real source] so it stays current.

Ask for it as one change, on top of a version that already works. Never as part of the first build.

Finished early? Don't polish — do this instead
  • Send it to one real person outside this room and ask what's confusing. Change the first thing they say.
  • Make a second, different one. The second takes a fraction of the time and teaches you more than polishing the first.
  • Go and sit with someone on yellow. Don't take their keyboard.
Tools down is [TIME] — earlier than you think, and it's deliberate. Publishing is the step that breaks under time pressure, so we stop while there's room to fix it.

Step 7

Go and look at what everyone built

30 minMove around

No presentations, no stage, no slides. Everyone's link goes on one page, and we walk around and look.

How it works

  • Add your link to [GALLERY PAGE].
  • Stay near your laptop for the first half. People will come to you.
  • When someone arrives, you get about a minute — what it is, who it's for, one thing that surprised you.
  • Swap halfway and go and look at other people's.
Why it's set up this way: you'll explain yours several times to two or three people rather than once to forty. It gets sharper each time, and nobody has to stand at the front. That's a deliberate choice, not a scheduling compromise.

Worth asking people

  • "What did it get wrong first?" — the most useful question in the room
  • "What did you cut?"
  • "What would you do with another hour?"

Step 8 · close

What you did and didn't just learn

10 min

Being straight with you, because the honest version is more useful than the flattering one.

What you can now do

  • Get an idea out of your head and onto a screen fast enough to react to it
  • Improve something by describing what's wrong, one change at a time
  • Put something on the internet without asking anyone
  • Point the thing at your real work instead of a blank page
  • Notice when it's confidently wrong — the skill that makes the rest safe

What you haven't learned

You haven't learned to build software, and you don't need to have. What you built today works; you couldn't yet fix it if it broke in a way the model can't talk its way out of.

The practical line: build anything you like for yourself and your team. Before something depends on it — money, customer data, a decision someone else makes — bring an engineer in. Not because you did it wrong, but because "who fixes this at 2am" is a real question and today didn't answer it.

The thing that actually predicts whether today mattered is not what's on your screen right now. It's whether you build a second thing in the next fortnight that nobody asked you to build. That's the whole difference between a good day out and a change in how you work.

Where to go

  • This site stays up. Every step is here, in order, whenever you need it.
  • Your page stays live. Keep changing it.
  • [SUPPORT CHANNEL] — where to ask when something breaks, and it will.