Want to build with us?
Build the thing.
We're Forward Deployed Engineers — we don't deliver slide decks, we ship production AI systems. This is the exact track we'd hand a new engineer: five phases where you build and deploy a real AI system on cloud infrastructure you own, by directing an AI assistant the way we work every day.
It's interactive and self-paced. Copy the prompts, check off your progress, pass each phase's knowledge check. Finish it and you'll have a live, deployed AI application — and a reason to reach out.
Progress is saved to this browser only — no sign-in needed. Start with Phase 0.
How to Learn This — By Asking AI
This is a five-part, hands-on track that takes you from “I can write code but I've never really built with AI” to “I've built and deployed a full AI system on cloud infrastructure I own.” You build one project — a personal portfolio site on Google Cloud — and each phase adds a new capability to it.
You are not expected to already know this material, and you are not expected to memorize commands. Throughout every phase you'll see “▸ Ask Claude: …” prompts — hand those to your AI tool. It does the typing; you direct, review, and learn. The real skill isn't typing cloud commands. It's learning to work with an AI over many rounds to build something genuinely good — and to stop and ask whenever something doesn't make sense.
The loop you'll repeat constantly
Describe what you want
In plain English. “Set up X” or “Explain how Y works in this project.”
Let the AI draft it
Code, a config file, or an explanation.
Review before you accept
Glance at what it actually did — which files, roughly what changed — rather than rubber-stamping.
Ask follow-ups until it's clear
“Why this way and not another?” “What does this line do?”
Ship it, then iterate
Nothing's right on the first try; refining is the job.
Prompts that get you unstuck
“I don't understand what a [term] is. Explain it like I'm new to cloud, with a small real-world example.”
Use this the instant a word is unfamiliar — repo, container, embedding, OAuth, whatever.
“Walk me through what this command / file does, line by line.”
Turns any code block in these docs into a lesson instead of a copy-paste.
“I got this error: [paste it]. What does it mean and how do I fix it?”
Paste the whole error. Errors are normal and are some of the best learning moments.
“Before you change anything, give me a plan for how you'd do this.”
“Plan Mode” thinking — catch a wrong assumption as a sentence, before it becomes five bad files.
“Why did you choose this approach over the alternatives?”
Builds real understanding of trade-offs — what separates engineers from code-pasters.
These prompts say “Claude,” but the workflow is the same with any capable agentic AI coding tool (Claude Code, Codex, and others). The key move is using a tool that can create files and run commands in your project — not just a chat window you copy-paste from. That's what lets you direct instead of type.
Five phases, one project
Each phase adds one capability to the same project and ends with something working you can show off. You can stop after any phase and still have a real, deployed result. Go in order — each phase assumes the last.
Foundations · Set Up Your Cloud Home Base
Build: The account, project, and guardrails. | Learn: Cloud basics, IAM, keyless auth (Workload Identity Federation).
Ship It · A Portfolio Site, Live on Cloud Run
Build: A live site, auto-deployed on PR merge. | Learn: Containers, Cloud Run, Terraform, Git & PRs, CI/CD.
Make It Smart · Add an LLM Feature with Vertex AI
Build: An LLM feature in the site. | Learn: Calling an LLM, prompting, Vertex AI.
Ground It · A RAG System with Cloud SQL + pgvector
Build: Answers from your own documents. | Learn: RAG, embeddings, vector search.
Open It Up · Connect Claude to Your Site with an OAuth MCP Server
Build: An AI connector to your site. | Learn: Agents, tools, MCP, OAuth.
AI systems aren't like normal software
Traditional code is deterministic: same input, same output. LLMs are not — the same prompt can give different answers, and a model can be confidently wrong. That's why testing, careful design, and evaluation aren't extras in AI engineering; they're the core of the job. If you come from normal software, this is the biggest mental adjustment.
Stay inside the free credit
Phase 0 has you claim Google Cloud's $300 free credit and set a budget alert before building anything. The whole track is designed to fit inside free/near-free usage. The one thing that bills continuously (rather than scaling to zero) is the database in Phase 3 — so that's the resource to shut down when you're not using it.
Worth bookmarking
Official hands-on guides for prompting, tool use, and RAG.
Get Claude Pro (referral)The plan that unlocks Claude Code, the tool that drives every phase.
Claude Code docsHow to use the agentic tool that drives every phase.
Google Cloud free programThe $300 trial and always-free tier you'll use throughout.
Built it? Show us.
If you work through this track and deploy your own AI system, that's the strongest application you could send. Reach out with a link to what you built and how you built it.