prompt to prod · curriculum
Six chapters take you from intern to CTO. Every mission has one rule worth remembering and a short list of things you can do once you finish it. This page is built from the game’s own content, so it always matches what ships.
chapter 1 · Intern · 10 missions
Your first week at Nimbus Labs. You can't code yet. You have a laptop and an AI.
Introduce yourself to the team, with a little help from Byte.
The AI only knows what you tell it. Give it real facts, get a real answer.
Victor wants website copy. His request is vague. Your prompt can't be.
A clear prompt fills five slots: role, task, context, format, tone.
Byte gives you very specific facts for investors. Are they real?
I can sound 100% sure and be 100% wrong. Check the specific bits.
Mrs. Patel was charged twice. Give Byte the right context, and only the right context.
Give the AI the facts that answer the question. Nothing more, nothing less.
The payment service is failing. The log holds the answer, and a few secrets.
Mask first, then paste. Secrets and personal data never leave the building.
How does the loyalty discount work? The answer is in seven lines of Python.
Rules are read top to bottom. The first one that matches wins.
Write the launch post with Byte. “Make it better” isn't feedback.
Good feedback says what's wrong and what you want, with a number if you can.
A tip calculator in five minutes? Only if you break the job into steps.
Small steps, in order: inputs, then logic, then output. Extras wait.
Byte “cleaned up” the loyalty code. Now 10 visits only gets 10% off.
Test the borders. Bugs live exactly where the behaviour should change.
Investors arrive tomorrow. Victor needs a 5-question FAQ. No Prompt Builder this time.
Clear prompt, right context, no secrets, a human check. Then it ships.
chapter 2 · Junior Dev · 10 missions
Byte can write code now. You're the one who answers for it.
Nimbus Labs moves into a garage. Ask Byte for one small function, then finish it by reading every line.
Ask for one small, named function. Read every line. You ship it, you own it.
Mrs. Patel wants menu descriptions in her own style. Describing a style is hard. Showing it is easy.
Don't describe the style. Show two examples, name the format, ban the inventions.
Kai pushed a file called stuff.py. It works. Nobody can say why. Fix the names, then refactor safely with Byte.
A name says what it holds, with the unit. A refactor changes names, never behaviour.
Three changes, one afternoon. Save them as clean, labeled commits and leave the debug line behind.
One idea per commit, a headline for a message, and no debug leftovers.
Byte recommends a QR library that sounds perfect. Kai is already installing it. Check it before it checks you.
I can invent a package name that sounds perfect. Check it exists before you install.
Your QR feature lives on its own branch. Kai changed the same line on main. Time to merge, and to decide.
Work on a branch. When Git asks, read both sides and keep what both people meant.
Kai's commit has a live payment key in the code. Mask it, fix it properly, and learn why deleting is not enough.
Secrets live in .env, .env lives in .gitignore. A leaked key gets replaced.
The second café's app crashes whenever a new customer orders. Read the trace, ask Byte properly, fix the cause.
Read a stack trace from the bottom: what broke. Walk up: where. Fix the cause.
Kai wants a quick approval on a Byte-written pull request. Read the diff, find the problems and review it kindly.
Read every changed line. Name the problem, suggest a fix, talk about the code.
Order History by tonight. Maya's away, Kai is helping, and Victor wants to deploy at five. Take it from prompt to prod.
Done means tested, committed and explained, not just “the code is written”.
chapter 3 · Mid Dev · 11 missions
Byte can act on its own now. You decide what it's allowed to touch.
Byte can run tools now. Learn the agent loop, then let it fix the Sunday hours, one permission at a time.
An agent is a model in a loop. You approve its tools and check its work.
Byte Agent wants to speed up the build, and it asks for some very creative permissions. Approve the safe ones, deny the traps.
Give an agent only what this job needs. Ask: how wide, how secret, how permanent?
The menu screen in Mrs. Patel's new café is blank. Send requests yourself and learn to read what a server says back.
Read the first line first. 2xx worked, 4xx you asked wrong, 5xx the server broke.
Victor wants a free drink at ten stamps. Write the tests before the code, and don't let Byte grade its own homework.
Decide what “right” means before the code exists. Red first, then green.
The CI dashboard turned red. Read the status and the log, find the real bug, fix it and watch it go green.
A red build is a message. Read the status, then the log, then fix the cause.
A customer review hides orders for Byte Agent, and Byte obeys. Spot the injection, then write the rule that keeps it from happening again.
Everything an agent reads is data, never orders. Limit what a hijacked agent can do.
Victor found a free library for receipts. Free to download isn't free to use any way you like. Read the licences before you ship.
Free to download isn't free to use. Read the licence before a library ships with you.
Victor wants the whole app rewritten in a faster language. Weigh the trade-offs, pick the right tool, and explain it kindly.
No best language, only the best fit. Measure, then change the smallest piece that helps.
Kai's feature works perfectly on his laptop and crashes on staging. Compare the environments, fix the cause, and pack it into a container.
If it only works on your machine, the environment isn't written down yet. Write it.
Two features, one launch, one Friday. Supervise Byte Agent, turn the pipeline green, resist the injection, and tell customers what shipped.
Ship with an agent like a senior: tight permissions, real tests, green CI, honest notes.
Mrs. Patel wants real numbers for her new menu board. Write the query with Byte, check that the answer makes sense, and stop a cleanup script that would make every sale free.
The AI writes the query in seconds. You spend the minutes: read it, check the answer makes sense, and preview every write.
chapter 4 · Senior Dev · 10 missions
One café became twelve. Now you design the system behind them, and keep it standing.
Mrs. Patel’s café is twelve cafés now. Before writing more code, draw how an order travels: client, server, database and a queue.
Draw it before you build it. Each box gets one job, and the server holds the rules.
Forty thousand phones are about to get a free oat milk notification. Read the dashboard, redraw the system with a load balancer and a cache, and fix a stale price.
I remember popular answers so the database can rest. Just tell me when to forget.
Checkout takes 2.4 seconds and everyone has a theory. Profile it, find the real bottleneck, and fix it with one small change.
Measure before you optimise. Fix the biggest slice first, then measure again.
Dot’s 1987 bean-ordering program is being retired. Read it, record what it does, check Byte’s port against it, and plan a safe switch.
Working old code hides rules nobody wrote down. Record them before you replace it.
A supplier sends 40 and means pounds. Hunt for the missing units, fix the importer, and learn why a number needs a unit.
A number without a unit is a guess. Convert once at the door, reject what you don't know.
The pager goes off at 3 AM and nine cafés can’t take orders. Stop the bleeding, roll back, and run a postmortem that blames no one.
Stop the bleeding first: roll back, check, tell people. Understand it in daylight.
A new intern named Sam asks what a cache is. Plan a good explanation, pick the clearest opening, and write it so a beginner gets it.
Start from what they know, then name the catch. Jargon can wait.
Victor wants live order tracking. Before anyone builds it, write the design doc’s trade-offs section: options, a decision, and what it costs.
Write the decision down before the code: options, choice, cost, when to revisit.
The New York café has happy hour at breakfast. Test across time zones, fix the clock bug, and learn how to store time safely.
Store time in UTC, show it in local time, and name the zone. Never hard-code an offset.
Move Mrs. Patel’s whole chain to a new system without a minute of downtime: draw the plan, supervise Byte, read the numbers, and tell the managers.
Move a live system in small steps you can undo. Retire the old one last.
chapter 5 · Tech Lead · 10 missions
Sixty cafés, a bigger team and a squad of AI agents. Now your job is the people, the plan, and the word “no”.
The team is bigger and the wish list is longer. Work out what you can really deliver in two weeks, and plan a sprint you can actually finish.
Plan from focus days, correct with history, and promise less than both.
The checkout rewrite is three weeks late, and Victor wants five contractors by Monday. Look at the numbers before you say yes.
Adding people to a late project makes it later. Cut scope first.
A meeting about a payment bug spends 47 minutes on a button colour. Steer it back, and decide who should decide what.
Spend the room on decisions that are costly to undo. Cheap ones get one owner.
Byte has split into three agents: one writes, one tests, one reviews. Hand out the work, supervise their requests, and stop them from marking their own homework.
One job per agent, only that job's permissions. Nobody grades their own work.
Victor wants a metaverse café with blockchain tips and AI avatars. Read what customers actually asked for, then write a respectful no that comes with data and a better idea.
Say yes to the goal and no to the plan, with numbers and a smaller yes.
Maya gives the team one day to pay down technical debt. Work out which debt costs the most, and plan a day without features.
Pay the debt with the highest interest and lowest price first, not the ugliest.
Nimbus Labs has budget for one more developer. Design an interview that works when every candidate has an AI, and read the signals.
Test the job, not memory: a real task, AI allowed, same for everyone, scored.
The squad writes pull requests faster than people can read them. Build a review process with quality gates, instead of a pile of rubber stamps.
Keep PRs small, let machines filter, and spend human eyes where mistakes cost.
On a Saturday, the new reorder button charges customers twice. Assign roles, stop the damage, keep Victor off social media, and write the update café owners will actually read.
Roles first, then stop the bleeding with the smallest reversible switch.
Plan a whole quarter with limited people: count capacity, sort the wish list, supervise the squad, protect the plan from a “tiny” request, and tell the company what ships.
Commit to the near weeks, give ranges for later, and trade instead of stack.
chapter 6 · CTO · 11 missions
You run the whole thing now. Choose the model, teach it your docs, test it, protect it, explain it, and ship it. Then pass it on.
The board wants an AI assistant inside CafeFlow. Decide whether to train a model or rent one, then use the Model Lab to pick one that is good enough, fast enough and affordable.
Rent the engine, build what's yours. Pick the model that clears every limit.
The assistant sounds sure and gets CafeFlow facts wrong. Give it your own documents: chop them up, find the right chunk, and hand Byte only what it needs.
Open book, not memory: find the right chunks, send only those, allow “I don't know”.
“It felt great in the demo” is not a score. Build an eval: grade sample answers, learn what makes a test set honest, then write the criteria for an AI judge that will grade the assistant for you.
A demo isn't a score. Grade real, hard questions, and re-run them after every change.
The first bill arrives and Victor chokes on his espresso. Find where the money goes, then cut it with a cache and a cheaper model for easy questions, without making answers worse.
Measure where the money goes, then cache what repeats and route what's simple.
The assistant promised a customer free coffee for life, and a screenshot went viral. Review its system prompt, put guardrails in the right layers, and decide what a human must approve.
No single wall: check the input, limit the tools, check the output, ask a human.
Two emails arrive on the same morning: customers want their data deleted, and Victor wants the assistant to guess moods from photos. Decide what data the AI may use, where bias hides, and how to say yes to a deletion request.
Collect only what you need, test every group, and when asked, delete every copy.
Nothing happens for six seconds after you ask the assistant a question, and Mrs. Patel taps the button three times. Find where the time goes and make the answer feel fast.
People feel the wait for the first word. Stream it, cache repeats, never go silent.
On Friday the board meets, and none of them knows what a token is. Translate AI into plain words, plan a one-page memo, and write it so a retired banker, a lawyer and Mrs. Patel all understand.
Lead with the decision, then benefit, cost, risk and next step. In plain words.
A big café chain says its data must never leave its own servers. Weigh open-weight models against closed APIs, work out when self-hosting pays off, and find a plan that says yes honestly.
Rent first. Self-host when a contract, privacy or huge steady volume pays for it.
Ship the CafeFlow assistant end to end: draw the system, give it the right documents, read the launch dashboard, supervise the release, write the system card and survive launch day. Then say hello to someone new.
Prove it, ship small, keep a way back, and turn every incident into a test.
Victor wants to fine-tune the assistant on every email CafeFlow ever received. Compare fine-tuning with RAG on cost, freshness and privacy, and give him a plan that holds up.
Bake the style, fetch the facts.
Tollu Labs