Wrap-up
You shipped an AI-built career website on Google Cloud. In about three hours you went from a blank project to a public URL — using the same toolchain professional teams ship with.
Pulse check — what stuck?
The same five questions you saw on the setup page. Answer each one out loud or in your head before clicking Show answer. Retrieval — pulling something from memory, not re-reading it — is the cheapest, best-supported memory tool there is (Adesope 2017 meta-analysis of 217 studies, Hedges' g = 0.61). Three minutes here will outlast the rest of the recap.
1. What's the role of the Gem versus the NotebookLM in step 3, and why split the work across two tools?
Show answer
The Gem runs the disciplined 8-question interview and outputs a YAML brief. NotebookLM takes the brief and generates the Stitch and Antigravity prompts — grounded in the workshop's research corpus, so it injects real regional employer names, salary anchors, and credential signals based on your Market Profile. Each tool does what it's best at: the Gem enforces interview discipline, the notebook gives citation-grounded output. A shared Gem can't have a NotebookLM corpus attached, so the split is also a sharing constraint, not just a quality choice.
2. What's persona.json used for, and why does it live as its own file instead of being baked into the HTML?
Show answer
It's the structured-data source of truth — name, projects, skills, contact, FAQ. The site renders from it at build time, and the bonus tracks (MCP server in step 8, chatbot in step 9) read from the same file. Same data, three surfaces. Editing one JSON file beats grepping HTML when content changes.
3. When nobody's hitting your deployed Cloud Run site, what does it cost?
Show answer
Effectively nothing. Cloud Run scales to zero when idle — no instance running means no compute time being billed. You pay per request (CPU + memory × seconds) only while traffic exists. That's why $5 of trial credit can host a portfolio site for months.
4. Where do your TryGCP credits live — on the project, or somewhere else?
Show answer
On the trial billing account, not on the project. The project consumes credits; the billing account holds them. One billing account can fund many projects. That's why the project dashboard doesn't show a credit balance — Billing → Credits is where it lives.
5. If you wanted to ship a different small site next week, which steps from today would you actually re-run?
Show answer
Spec (Gem interview → YAML brief → NotebookLM prompts) → Design (Stitch) → Build (Antigravity) → Deploy (gcloud run deploy --source .). Setup (step 1) is a one-time cost per Google account; you keep the cloud project, the credits, the gcloud install, and Antigravity. The rest is the reusable pipeline. That's the actual deliverable of today.
What you built
- A complete site spec — produced by a 10-minute Gemini Gem interview that output a YAML brief, then turned by NotebookLM (grounded in regional research) into a Stitch design prompt and an Antigravity build prompt.
- A visual design generated and refined in Stitch from that design prompt.
- A working career site coded agentically in Antigravity, including a
persona.jsondata file and the Stitch design as a copy-paste reference the agent matched against. - A live deployment on Cloud Run, sitting behind a public HTTPS URL.
- A real, professional development workflow: spec-first prompting, agent-driven scaffolding, small commits, iterative prompts.
Next steps
- Join the AI-native Discord. Free, ongoing, full of people who just shipped what you shipped. Drop your deployed URL, ask questions when you hit a roadblock, share what you build next. The single cheapest way to keep momentum after today.
- Customize further on your own time. Edit
persona.json(or re-run a focused Antigravity prompt), redeploy. The pipeline you set up today is reusable. - Try the bonus tracks. The MCP server turns your
persona.jsoninto tools any AI client can call. The chatbot wires Vertex AI (or your MCP server) to a chat widget on the deployed site. - Add the URL to LinkedIn, GitHub, and your CV. A live site that you can iterate on beats a static PDF every time. Put it where recruiters will see it.
- Explore the resources page. Curated links to Google products, frameworks, and further reading — bookmark it and come back as questions come up.
Keep going with codelabs
Right now is the highest-leverage moment to learn more — your Google Cloud project is configured, credits are live, and you have momentum. Google publishes free, hands-on tutorials called codelabs that fit cleanly into a focused hour. Three to consider next:
- Getting Started with Google Antigravity — deepen your fluency with the IDE you used today.
- Getting Started with Google MCP Servers — if the bonus track caught your attention.
- Building AI Agents with ADK: The Foundation — the natural next step from a chatbot to a real agent.
The full curated list with commentary lives on the resources page.
Five minutes of feedback
Before you close the tab — fill in the workshop feedback form. Five minutes, mostly tap-to-answer, one short text field. Honest answers (especially the "what didn't work" and "one thing you'd change" parts) directly shape the next run of this workshop.
What happens next
Most of what you saw today will fade unless you re-touch it. The forgetting curve is brutal but well-mapped — the same study you do in three blocks across a week beats the same study done all at once, by a wide margin. Cepeda et al. (2006) pooled 839 effect sizes across 317 experiments and found spacing wins essentially every time.
Three short retrieval prompts, scheduled across the next month. Each one takes under a minute. The point isn't to look anything up — it's to try to remember, even badly. Trying counts.
+ 24 hours — tomorrow, same time as the workshop
Without scrolling: what's the difference between a billing account and a project, and which one do credits live on? Peek at step 1 only after you've tried.
+ 7 days — one week out
Sketch the build-brief prompt shape from step 5 §6 on paper, from memory, in 60 seconds. If you can reproduce it — its inputs (the design reference, the brief), outputs (named files, a launch checklist), scope (the section list), constraints (plain HTML/CSS), and verification (refresh localhost) — the pattern is yours.
+ 30 days — one month out
Did you ship a second site using the Gem → Stitch → Antigravity → Cloud Run pipeline? If yes, the workshop landed. If no — what's stopping you? (The honest answer is usually "I haven't tried" — fixing that is one evening.)
The cheapest way to actually do this: open the email below in your client and use scheduled-send for tomorrow morning, +7 days, and +30 days. Three drafts, three minutes.
Open three retrieval drafts in my email
Resources
Glossary
Short definitions for every term used in the workshop.
Resources
Google products, frameworks, slides, and further reading.
About the trainer
Who ran today's workshop and how to stay in touch.
You just shipped a real product on the exact stack — Antigravity, Gemini, Stitch, Cloud Run — that most professional developers haven't touched. The hard part is over.
What keeps you here is what you do next. Use it or lose it — run an agent on something real every week, even something small. Stay current — pick one channel (YouTube, newsletter, podcast) and check it weekly. Build in public — post the repos, write what worked. The market in 2026 is brutal; the engineers who do the reps are the ones still working in 2027.
If you haven't yet — drop your deployed URL into the workshop channel and click on a couple of your neighbours' sites. Today's pipeline produces about thirty live URLs in this room; that's its own small graduating class.