Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
Vibe Coding for Beginners (Full Course 2026)
TL;DR
Riley’s core pitch is that beginners can now ship real software with mostly prompts — using Codex with GPT-5.5, he goes from zero-code explanations to a deployed multi-user app, then turns it into a desktop Electron app and an iOS Swift app in one 55-minute course.
He makes the mental model dead simple: an app is basically a folder of files AI edits for you — his first demo is a Microsoft Paint-style React/Vite web app built in one prompt, then restyled “like Apple” with undo support after a 5 minute 44 second generation run.
GitHub gets reframed as “Google Drive for nerds,” not a scary developer tool — Riley shows how Codex can create, update, and commit code to a private repo, with commits acting like saved versions of the app.
The real leap happens when he adds Firebase for auth, database, and storage — after a 2-5 minute setup, the ‘Shared Brain’ app gets Google Sign-In, Firestore-backed saved items, file uploads, and a shared team workspace that can ingest X, YouTube, Instagram, and screenshots.
The workflow is iterative, not magical: prompt, test, inspect, paste errors, repeat — he fixes permission issues, metadata bugs, storage authorization, and UI rough edges by feeding Codex console logs, screenshots, and voice notes through WhisperFlow.
The most striking demo is cross-platform reuse of one backend — after the web app works on Vercel, Codex creates an Electron desktop version and an iOS simulator build that both sync to the same Firebase backend, with Riley calling the result “mind-blowing” and “98% of the work done in a single prompt.”
The Breakdown
Codex, GPT-5.5, and the beginner-friendly mental model
Riley opens by framing vibe coding as one of the fastest-moving parts of AI in 2026, and says the trick is to stop trying to absorb everything and just focus for a few hours. His big simplification: an app is basically a folder of code files, and Codex plus GPT-5.5 is now good enough that beginners can have AI create and edit those files for them.
The one-prompt Paint app that makes the whole thing click
To make it concrete, he creates a simple Microsoft Paint-style web app in Codex and waits 5 minutes and 44 seconds while GPT-5.5 builds it as a React/Vite project. Then he tweaks it with plain English — “make this look like Apple” and add undo — and the app updates live, which is his way of showing the core loop: prompt, preview locally, prompt again.
GitHub becomes “Google Drive for nerds”
He pauses to demystify saving work, especially for non-coders who find GitHub intimidating. Riley’s framing is memorable and useful: a local folder is a directory, the cloud version is a repository, and commits are basically “please update the saved version” — which lets Codex push the app into a private GitHub repo and keep versions cleanly synced.
Building “Shared Brain,” a premium internal second-brain app
The main project is an internal tool called Shared Brain: a visual library for video ideas, links, screenshots, hooks, references, and loose creative thoughts, inspired by how much he likes the app MyMind. He wants it to feel “like my mind,” not like a boring admin dashboard, and he also wants agents to be able to read from it and add to it, complete with copyable instructions and credentials inside the user profile.
Firebase setup: the only real off-platform step
Before asking Codex to build the real app, he sets up Firebase for the three things the Paint demo lacked: authentication, database, and storage. He walks through creating a Firebase project, enabling Firestore, upgrading for storage with a capped budget like $20, turning on Google Sign-In, and carefully warning viewers not to expose the Firebase config because it functions like a password.
Debugging in public: permission errors, console logs, and messy metadata
The first full build takes 25 minutes and 51 seconds, and while the app mostly works, Riley immediately hits classic real-world issues: popup/auth problems, insufficient permissions, and broken saves. Instead of pretending it’s magic, he right-clicks Inspect, copies console errors like 403 responses back into Codex, and keeps iterating until X posts, YouTube links, Instagram posts, and screenshot uploads all save correctly.
Polishing the product with AI titles, better cards, and team filters
Once the plumbing works, the session becomes product design by prompt. He adds lightweight OpenAI API-based title generation using 4.1 nano, improves thumbnails and card layouts, switches to masonry-style grids, and uses screenshots plus WhisperFlow voice notes to refine filters for category, platform, and user so the app works as a true multiplayer internal tool.
From localhost to Vercel to Electron and iPhone simulator
After deploying the web app to Vercel, he fixes Firebase’s authorized domains so Google auth works on the public URL, then asks Codex for a plan to turn the same project into a desktop and iOS app. In about 29 minutes and 52 seconds it creates an Electron desktop app and a Swift/Xcode iOS build; after a couple more fixes for popup auth and a keychain error, all three clients sync to the same backend, and Riley’s reaction is basically disbelief that this can now be done in “less than seven prompts.”
The ending bet: this is enough to start building for real
He closes by saying this is already enough for most people to get their feet wet with web, desktop, or iOS app building, even if deeper App Store and distribution details deserve separate videos. The final proof point is the agent skill itself: he pastes the skill into a new chat, has the agent add tweets into Shared Brain on his behalf, checks the app, and sees the posts appear.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
