Step 01 · Sense
Make something small and look at it
Before install and syntax: write a request → see the screen → ask for a change.
Output: a one-page web app you can deploy
Tools and mistakes from actual Giant Stream work, arranged as news, a glossary, a path, and practice. The aim is that a first-time visitor can finish one small result.
Rather than listing every new model, we look at what actually changes in the work. Search results sit next to Giant Stream notes, with a link to the original.
Tooltips help while you read. The glossary is for looking things up. Giant Stream plain-word entries come first; other terms come from the Stack Overflow tag wiki.
Each step is linear. Goals, tools, 8–10 lessons, then a capstone. The next step stays closed until that capstone is done.
Before install and syntax: write a request → see the screen → ask for a change.
Do not stop at a screen. Turn the idea into a small product with login, a database, an API, and a deploy.
Use AGENTS.md, MCP, and an eval loop to make a repeatable work system, then apply it to a team or a business.
Goal, limits, checks, and a follow-up question on one card, reused when the tool changes.
What you asked, what broke, and how you fixed it.
Env vars, domain, rollback, mobile layout, and security notes before a release.
Capstones from each step, grouped as public work for a conversation, a hire, or a sale.
Instead of typing code, write intent in plain language and trim the result. No install. Browser only. Aim: one working web app in the first week.
Beginner stays out of an IDE. Chat and an app builder only.
Not “writing code” so much as stating intent — how that differs from ordinary coding.
20 min · ideaOne line to Lovable: “a Giant Stream-tone landing page” → a working page in a few minutes.
25 min · practiceSix signs a result is off, and a follow-up prompt to fix it.
30 min · sense“Do not ask for everything at once” — the first of five working rules.
25 min · patternA chatbot with a voice, made in Bolt from plain language only.
40 min · projectA small tool that takes a URL and summarizes it — a first feel for data flow.
45 min · projectA screenshot or one sentence in; a UI component out.
35 min · UIWhat / for whom / how, in one sentence.
30 min · capstone prepAny topic, as long as you would actually use it each week. Share a deployed URL to finish beginner.
Move the beginner loop onto an IDE agent (Cursor · Claude Code · Windsurf), then run one cycle: front, back, database, payment, deploy. The five working rules should become habit here.
From intermediate on, the work happens in a terminal and a VS Code-class agent.
Limits of an app builder, why codebase context matters, what a multi-file patch can do.
25 min · ideaStrengths, gaps, and when to use which. Pick one.
35 min · toolsThe model only knows what you showed it. Attach just enough to narrow the job.
30 min · patternFigma screenshot → v0 component → Cursor folds it into the project.
45 min · practiceSupabase + tRPC or Hono in the same loop. Schema → handler → checks.
50 min · full stack“Add login / add payment” is not the whole job. Cover the leftover parts.
50 min · full stackDo not generate code alone. Keep a small eval loop that catches regressions.
40 min · qualityA deploy that ends at git push, including env vars, preview, and rollback.
35 min · ops“Fix it” is slower than “give me three hypotheses for this behavior.”
35 min · debugContext · why · chain of thought · role · small repeats — a self-check.
30 min · wrapA one-person SaaS with a real payment path. Domain, TLS, pay, and cancel, all in the same loop. After a week of running it, write a short note on traffic and errors.
Past one-off vibe coding. Treat the model as a regular part of an engineering process. AGENTS.md for how the agent should behave, MCP for live data, an eval loop and more than one agent so results stay even.
Advanced also covers an agent directing another agent.
The point where ad-hoc work has to become a system. What has to change.
30 min · stanceConventions, stack, and bans in one document the agent can read.
40 min · docsHow an agent stays oriented in a large codebase.
35 min · docsConnect Slack, GitHub, or your own database as an MCP server so the agent sees live context.
50 min · infraExpose an internal API over MCP so the agent knows the company domain.
60 min · practiceSplit designer, implementer, and reviewer. Let them check each other.
55 min · designBuild an eval set and catch regressions every day.
50 min · qualityA checklist for the quiet gaps in code that was written quickly.
45 min · opsHow a person reviews a PR the agent wrote. Ownership and agreement, restated.
45 min · teamWorking rules that stay put while tools and models move.
30 min · outlookAutomate one real task (for example: classify a customer note, draft a reply, write it into an internal system). Submit AGENTS.md, MCP, and an eval loop with it.
Paid courses now sell projects, a completion record, a security check, and a portfolio more than an intro lecture. This hub does not copy that shape. Beta work is arranged so a participant leaves something they can run.
Prompt cards for a landing page, chatbot, summary tool, and an internal tool, plus a recovery prompt when it fails.
A 30-minute task each day. In the last week, submit one public link to your own service.
AGENTS.md, a work brief, review prompts, and branch rules in one agent setup.
A one-time review of env vars, public keys, a payment test, permissions, rollback, and mobile layout.
The playground is not a large IDE. It is a room of small product templates Giant Stream has already used, so you can follow the prompt, the code, and the deploy.
// you fill this in goal: "a small brand landing page" tone: "calm, technical" required: about, products, contact // Giant Stream template adds opening line, CTA, card structure, mobile checklist
News and beginner notes stay public. The full path, working templates, and feedback need ongoing work, so those go into a paid beta first.
No payment. Public news, the glossary, and sample lessons.
A paid beta for people who want to actually build. Price and scope can move with beta feedback.
You write intent in plain language and the model writes the code and the result. The work is the loop: look, trim, adjust the intent. This hub is built so that loop can stay on one screen.
Yes. Track 01 is for someone who has never written a line. It starts with writing intent, reading a result, and noticing a miss.
Not yet. A Pro Beta around ₩19,900/month is the current test price. Waitlist and training/collaboration notes come first.
One model call when the playground writes or edits code from plain language. The credit policy is not fixed; it waits on beta cost data.
You do. Giant Stream only hosts and shares the work. Code, design, and content stay with the author.
Yes, it is in beta. The news board is being automated in steps; the playground opens in pieces. Beta participants will hear separately when it opens fully.