Courses/AI Kitchen/Episode 2
The Kitchen Brigade
One chef isn't enough. Lan stops cooking and starts managing: tasks, plans, a brigade of AI agents with roles, pipelines and a router.
Orders flood in after opening day. Lan wants a loyalty app and a Tết menu at the same time, and falls asleep typing to one chef at 1 a.m. Tú teaches her to turn one big wish into small tickets, ask for a plan before cooking, run chefs side by side, give them roles, put them on a conveyor belt, let the app pick the right chef, and taste on a real phone.
22 new words, from Task to Overhead. Honest note: a team costs more quota. Small fixes need one chef, not five.
What you will learn
- Turn one big idea into small tasks with clear "done means…" criteria
- Ask an AI agent for a plan before it builds, and approve it
- Run several AI agents side by side and send one message to all of them
- Give agents roles: a lead, builders and a reviewer (agent orchestration)
- Pipelines (plan → build → review) and a router that picks the right agent
- Test on real phone sizes, and know when one agent is cheaper than a whole team
- Task
- one work ticket
- Board
- the rail of work tickets
- Spec
- the standard recipe
- Acceptance criteria
- what done means
- Parallel
- several chefs working at once
- Agent team
- a team of AI chefs
- Send to many
- one message to the whole team
- Orchestration
- coordinating the whole team
- Subagent
- a line cook taking orders
- Reviewer
- the taster who checks the dish
- Role
- each member's job
- Pipeline
- a conveyor belt of stations
- Stage
- one station on the belt
- Approval
- the owner signs off first
- Router
- the job board
- Fallback
- the backup chef
- Default
- ready-made rules
- Responsive
- fits every screen size
- Emulator
- a pretend phone for testing
- Device
- a real piece of hardware
- Token
- the unit AI is billed in
- Overhead
- the cost of coordinating
Task: Split your big idea into 5 small tickets, each with a line saying what “done” means.
Split this idea into small tasks, each with clear acceptance criteria: [your idea]
Task: Open two agents in one project, give each a different task, then send them one shared message with Send to many.
Task: Draw a 3-role team in the Builder: a lead agent, a worker agent, and a reviewer agent.
Task: Create a 3-stage pipeline: plan → build → review, with approval on the plan stage.
Task: Open the Router, pick the defaults, then add a rule: design work → your favorite agent.
Task: Open the Mobile screen, view your page at phone size, and ask an agent to fix whatever breaks.
Task: Before your next job, ask: small, medium or big? Then pick the number of chefs to match.
Generated from the episode’s subtitles. Click a line to jump there.
Tu: Last time, Lan built an ordering website for Cloud Café, just by talking to one AI chef.
Lan: And I learned twenty-eight new words!
Lan: Since opening day, the orders haven't stopped. And every customer asks: do you have a loyalty card? Then there's the Tết menu page, and the second branch. I type to the chef, wait, type again, and wait again. It's one in the morning...
Tu: Lan! Did you fall asleep on the keyboard?
Lan: The chef is fast. But everything has to go through me, so I'm the slowest one in the kitchen.
Tu: Because you're still the cook. Starting today, you need to be the manager.
Tu: In this episode, we learn how to run a whole team of AI chefs. Eight chapters, from a ticket board all the way to a loyalty card running on a real phone. And the vocabulary bar picks up from last episode's twenty-eight words.
Tu: Review from last time: what's a prompt?
Lan: An order ticket! I haven't forgotten.
Lan: I want a loyalty app. Can I just type one sentence: make a loyalty app for the café?
Tu: You can, but that's way too big for one ticket. A big restaurant doesn't write 'cook a wedding banquet' on one slip. They split it up: starters, mains, desserts.
Tu: Each small job is called a task. A work ticket. All the tickets hang on a long rail in the kitchen. Whoever is free takes the next one. That's the board. And a spec is the standard recipe: what the dish needs and how it should look.
Lan: So how many tickets does the loyalty app become?
Tu: Let's try. One: the loyalty card page. Two: saving each customer's stamps. Three: the buy ten, get one free rule. Four: a QR code to scan at the counter. Five: testing on phones. And six: the Tết menu page, as its own ticket.
Lan: Six small tickets feel so much lighter than one giant dish.
Tu: 1DevTool has a Tasks board. Each ticket is a row, with a name, a description, and a status: to do, in progress, or done. The AI chefs can read this board. They pick up tickets, mark them done, and report back to you.
Lan: Just like the ticket rail in a real kitchen!
Lan: So how do I write a good ticket?
Tu: Look at this one: make a loyalty app, make it nice and easy to use. The chef would have ten questions. Nice how? Easy for who? A good ticket has three parts: a short name, a description of why it's needed, and a list of what done means.
Tu: The secret is the last part of every ticket: what done means. In the industry it's called acceptance criteria. For example, the buy ten get one free ticket. Done means: with ten stamps, the eleventh cup is free, and after redeeming,
Tu: the card goes back to zero.
Lan: With that line, the chef can't argue about whether it's finished.
Tu: Exactly. And neither can you.
Tu: If your ticket is a bit messy, just ask the AI to rewrite it. It adds what's missing, and you just read and approve.
Lan: I write it badly, the AI rewrites it. So handy.
Lan: But who writes the tickets? Me or the chef?
Tu: Either. You can write them yourself. Or describe the big idea and ask the head chef to split it into tickets, then you read and fix them. But whoever writes them, you approve them. A wrong ticket means the whole team builds the wrong thing,
Tu: no matter how fast.
Tu: New words in this chapter: task, board, spec, and acceptance criteria.
Tu: Try it now: take your biggest idea, split it into five small tickets, and write what done means for each one.
Chapter 2: Two kitchens, one call
Tu: Quick review: what's a task?
Lan: A work ticket on the rail. And every ticket says what done means.
Lan: I have six tickets now. But the chef only does one at a time. It's too slow.
Tu: Then bring in more chefs. Last time you invited a second chef to check the food. Now let two of them cook at the same time, one ticket each.
Tu: Two people working at the same time is called parallel. One chef builds the loyalty card page, the other builds the Tết menu page. Nobody waits for anybody.
Tu: In 1DevTool, each chef gets its own window. One glance and you see who's cooking what.
Lan: This one is doing the loyalty card, that one is doing the Tết page.
Tu: And when a chef needs you, its window lets you know.
Lan: What if both chefs ask me something at the same time?
Tu: Then you answer them one by one, like two cooks calling the owner. Go to whichever window says needs you first. A little tip: give each chef tickets that don't overlap. One on the loyalty card, one on the menu. What happens when they do overlap?
Tu: We'll see in the next chapter.
Lan: Wait. I want the whole café in red and gold for Tết. Do I have to type the same sentence to every chef?
Tu: No need. There's Send to many: write it once, send it to as many chefs as you like.
Tu: Pick the chefs you want, write one message, and press send. Each one gets exactly that message in its own window.
Lan: Like the head chef calling out: everyone, switch to the Tết plates! And everyone hears.
Tu: New words: parallel, agent team, and send to many.
Tu: Try it now: open two chefs in the same project, give each one a different ticket, then send them both one shared message.
Chapter 3: A brigade with roles
Lan: Two chefs are definitely faster. But yesterday they both edited the same file and stepped on each other's toes.
Tu: Because the team has no roles yet. A real kitchen always has a head chef, line cooks, and a taster. Everyone knows their job.
Tu: The head chef takes your request, splits it into tickets, and hands them to the line cooks. The line cooks do the cooking. Each takes a ticket, finishes it, and hands it up. The taster cooks nothing at all.
Tu: Their only job is to check every dish before it reaches you.
Lan: So I only have to talk to the head chef?
Tu: Exactly. You talk to one person, and the whole team gets to work.
Tu: The head chef splitting up and coordinating the work is called orchestration. A line cook taking work from the head chef is called a subagent. And the taster is called a reviewer.
Tu: In 1DevTool, you draw your kitchen team with the Builder. Each box is a chef, and each line shows who hands work to whom. Just drag and drop, no commands. Draw it once and save it. Next time, the whole team is in position.
Lan: Like the shift chart pinned up in the kitchen.
AI Head Chef: Hi Lan, I'm the head chef. I've split the loyalty app into six tickets and given them to two line cooks. Our taster will check every dish.
AI Taster: I'm the taster. I just found a bug: when a customer has exactly ten stamps, the free cup is counted twice. I've sent the ticket back to the cook.
Lan: The bug was found before it even reached me! This team is good.
Lan: What does the taster actually check?
Tu: Three questions. Does it match the ticket, meaning every line of what done means? Does it work? Click it, type in it, try to break it. And will customers understand it?
Lan: Why not let the chef who cooked it taste it? It'd be faster.
Tu: Because cooks usually think their own dish is delicious. Fresh eyes catch what the cook missed. True for real kitchens, and true for people too.
Tu: New words: orchestration, subagent, reviewer, and role.
Tu: Try it now: draw a team of three in the Builder. A head chef, a line cook, and a taster.
Chapter 4: The conveyor belt
Tu: Pop quiz: who on the team only checks, and never cooks?
Lan: The taster. The reviewer.
Lan: The team has roles now. But I want every dish to go in the right order: plan, cook, taste, and only then serve.
Tu: Then let's build a conveyor belt. The dish moves station by station, and only moves on when a station is done.
Tu: Station one: plan. Station two: cook. Station three: taste. Station four: test run. Each station has its own chef. If a station finds a problem, the dish is sent back and doesn't move on.
Lan: Like a mooncake packing line!
Lan: What if the tasting station finds a bad dish? Do we throw it all out?
Tu: No. The dish goes back to the cooking station with the taster's notes: what's wrong and what to fix. The cook fixes it, and the dish moves on again. That loop can happen a few times. But by the time it reaches you, it has passed every station.
Tu: The conveyor belt is called a pipeline. Each station is a stage. And you can add an approval gate: the plan has to be approved by you before any cooking starts.
Tu: This is a pipeline in 1DevTool. The stations are in order, and each station's result passes automatically to the next.
Lan: No more carrying dishes from one person to the next myself.
AI Head Chef: Lan, here's the plan for the QR code ticket: a unique code for every customer, and scanning it at the counter adds one stamp. Do you approve?
Lan: Approved. But add this: one customer can get at most three stamps a day.
AI Head Chef: Done, I've added that rule to the plan.
Tu: See? Fixing the plan took ten seconds. Fixing it after cooking would take all afternoon.
Tu: New words: pipeline, stage, and approval.
Tu: Try it now: build a three-station belt, plan, build, check, and put an approval gate on the plan station.
Chapter 5: Who cooks what
Lan: Now there are five chefs, and I keep having to remember: design work goes to this one, bug fixes go to that one. It's confusing.
Tu: Then let a job board do it for you. In 1DevTool, it's called the Router.
Tu: The Router is like the board at the kitchen door: grill orders go to the grill station, soup orders go to the soup station. For example: design work goes to the chef who's great at design.
Tu: Data work goes to the careful chef. Small chores go to the fast, cheap chef.
Lan: So I just drop tickets in, and the board sorts them.
Lan: Let's try it. Three new tickets: change the order button color, add a birthday column for customers, and fix a typo on the Tết page.
AI Head Chef: Got it. Following the job board: the button color goes to the design chef. The birthday column goes to the careful chef. And the typo goes to the fast chef.
AI Line Cook: I'll take the typo. Done, it took one minute.
Tu: This is the real Router table. Each row is a rule: which kind of work goes to which chef. You don't have to write the rules from scratch. Start from the defaults, then adjust them to fit your café.
Tu: One more thing: fallback. A backup chef. If the main chef runs out of usage halfway through, the ticket moves to the backup automatically. The kitchen never has to close.
Lan: If I'd had this before, I wouldn't have been stuck waiting.
AI Head Chef: Lan, I've used up my quota for this shift. I'm only halfway through the QR code ticket.
AI Line Cook: I'm the backup chef. I've read what was half done in the project. I'll finish the rest.
Lan: I didn't have to do a thing. The kitchen changed shifts by itself.
Tu: New words: router, fallback, and default.
Tu: Try it now: open the Router, pick the default rules, then add one rule of your own.
Chapter 6: Tasting on a phone
Tu: Quick one: what is the router like?
Lan: The job board at the kitchen door.
AI Taster: Lan, I've tasted everything on the computer, and it's all fine. But our customers use the loyalty card on their phones.
Lan: Oh, right. It looks great on a big screen. That doesn't mean it looks great on a phone.
Tu: A website that stretches to fit any screen is called responsive. Like a dish you can serve on a big plate or a small one. There are two ways to test it. One: a pretend phone running right on your computer, called an emulator. Two: plug in a real phone.
Tu: 1DevTool has a Mobile screen. You immediately see the loyalty card on a phone, at its real size. And if you plug in a real Android phone with a cable, 1DevTool picks it up, so you can tap through it with your own fingers.
Lan: Seven stamps filled in, three more to a free cup. It's beautiful!
Tu: One more thing: no two customers use the same phone. Some have old phones with small screens, some have new phones with tall screens, and the counter uses a tablet. Test at least those three sizes.
Lan: Then I'll have the taster test all three, and write a ticket for anything that breaks.
Lan: Mochi! Stop pressing my phone!
AI Taster: Wait, Lan. The cat just pressed the scan button ten times in a row, and the card added ten stamps. That's a real bug.
Lan: Mochi is the best tester in the café. New ticket: block rapid taps, max three stamps a day. Send it to the head chef!
Tu: New words: responsive, emulator, and device.
Tu: Try it now: open the Mobile screen, look at your page at phone size, and ask a chef to fix anything that breaks.
Chapter 7: When you don't need the whole team
Lan: Having a team is wonderful. From now on, I'll give every job to the whole team.
Tu: Hold on. I have to be honest with you in this chapter: not every job needs the whole team.
Tu: Every chef that works uses up quota. Five chefs means several times the usage. Calling the whole team to fix one typo is like opening a wedding banquet kitchen just to boil an egg.
Lan: So when do I call the whole team?
Tu: I have a simple rule. Small job: one chef. Medium job: two chefs and a taster. Big job, like the whole loyalty app: the full team and the conveyor belt.
Tu: And while the whole team is working, look at the top bar. The words need you show how many jobs are waiting for you. Click it to see what needs you first. No need to check every window.
Lan: Managing without running around. I like it.
Lan: Honest question: does a kitchen team cost a lot more?
Tu: It depends on the job. Small jobs with one chef cost about the same as last episode. Big jobs with the whole team cost more, but finish much faster.
Lan: Will several chefs slow down the café's computer?
Tu: The heavy lifting happens on the AI company's servers. Your computer mostly just shows text, so a normal office PC can run several chefs.
Lan: Last one: can I mix cheap chefs and expensive chefs?
Tu: Yes, and you should. With the Router, small jobs go to cheap chefs and hard jobs go to the best ones. Fast and economical.
Tu: One more common word: token. It's the unit AI companies use to measure usage and charge you. The bigger the team, the more tokens. And the effort spent just splitting up and coordinating work is called overhead. The smaller the job,
Tu: the more likely the overhead is bigger than the job itself.
Tu: New words: token, and overhead.
Tu: Try it now: next time before you hand out a job, ask yourself whether it's small, medium or big, and pick the number of chefs to match.
Chapter 8: Launching the loyalty card
AI Head Chef: Lan, all six tickets are done. The taster has checked everything, and the belt has run through every station.
AI Taster: I've tested it on the computer, on the pretend phone, and on a real phone. The cat's bug is fixed too.
Lan: Then let's launch!
Tu: Before we launch, let's fast-forward through the whole journey. One big dish became six tickets. The head chef split up the work, and the router sent it to the right chefs. Two line cooks cooked in parallel.
Tu: Every dish went down the belt: a plan you approved, cooking, tasting, and a test run. The taster tried three screen sizes, and a real phone.
Lan: And the cat found a bug. Six tickets, all done.
Lan: Our first customer to use the card is Minh, a regular who comes in every morning. The tenth stamp! Minh gets a free white coffee!
Tu: White coffee again.
Tu: Let's look back at the whole episode. The kitchen team's golden rule has four steps. Last episode we had the golden loop for one chef. This is the version for a whole team. One: split it up. Two: assign it. Three: taste it carefully.
Tu: Four: merge it and launch.
Lan: Split it, assign it, taste it, merge it. Got it.
Tu: This episode you learned task, board, spec, and acceptance criteria. You learned parallel and send to many. You learned orchestration, subagent and reviewer. You learned pipelines, with stages and approval gates.
Tu: Then router and fallback, responsive and emulator, and finally token and overhead.
Lan: Twenty-two more words. Fifty in total! I leveled up!
Tu: Did you notice? In this whole episode, you didn't cook a single dish. You split up the work, approved the plans, and tasted the final dish.
Lan: I really am a manager now.
Lan: Hello? Bin? You want to help out in the kitchen? Sure, come over this weekend.
Tu: Next time: The Shared Kitchen. When real people and AI chefs cook in the same pot.
Tu: Download 1DevTool, build your first kitchen team, and hand out your first work ticket today.
Lan: See you at Cloud Café, in episode three!
AI Taster: I've tasted every dish. And I think I need to taste one more white coffee, just to be sure.
Lan: What kind of taster keeps on tasting!