Beyond Vibe Coding: Building High-Craft, Production Apps with AI
Follow one service-status app from a vague brief to a verified release. Learn to guide coding agents, design interfaces people can use, connect real data, and decide when a change is ready to ship.
Narrated by the SanX AI Instructor
Free to take. Best if you already know basic TypeScript, Git, and how a web app stores data.
Read the intro transcript
A green toast can lieAn operator posts an incident update. The screen says published. On the public status page, nothing changes. So what did the agent actually build? Welcome to Beyond Vibe Coding, from SanX Academy. I’m your SanX AI Instructor. We'll follow that one update from the first request to a release someone can trust.
Give the agent a better targetWe'll pin down who can publish, where the update should appear, and how to tell that it really made it there. Then we'll tune the status page and its motion, use React Grab to find the exact badge, and build the save behind the button. Each tool earns its place by solving a problem you can see.
Make the release callWe'll watch the update reach the public page in a real browser. You'll see a separate access check and decide whether the evidence is enough to ship. All sixteen lessons and the scenario quizzes are free. Start with the question that makes every later check possible. What exactly should change?
How Experienced Engineers Build with Coding Agents
In a real business, software isn't built by typing vague wishes into a chat box. We anchor this entire course in one consistent, high-stakes application: a Service Status and Incident Management Dashboard.
High-Stakes Reality
Your payment gateway or messaging pipeline slows down. Thousands of users hit your status page. An authorized operator needs to publish an update immediately—without leaking private internal logs, breaking on mobile, or failing silently.
Plausible Illusions
Naive AI prompting outputs convincing prototypes that look ready but fail in production: buttons that trigger fake timers, hallucinated database dependencies, low-contrast text, missing access controls, and unverified releases.
Supervised Discipline
You direct the AI through a structured engineering loop: reverse interviewing to extract requirements, persistent repository rules, semantic design tokens, verified database migrations, and automated headless browser tests.
Course roadmap
5 modules · 16 lessons · 5 scenario quizzes
Phase 1Modules 01–02
Prepare the agent
- Outcome
- Scope the incident update and set up an agent you can supervise.
01Before You Prompt: Architecture, Project Memory, and Current Docs3 lessons
Module 01 · Real-World ScenarioNever ask an AI to write code until you have defined what success and failure look like.
Imagine your payment or messaging service degrades during peak hours. You need your team to publish an urgent incident notice to customers. If you ask an AI coding agent with a vague prompt like "build a status page with a publish button", it quickly outputs a flashy mockup with billing tabs and animations you never requested, but clicking Publish does not actually save or update anything.
⚠️ The Vibe Coding TrapTyping a one-sentence wish into an AI chat and assuming the resulting code is complete because the user interface looks convincing.
✓ The Engineering SolutionRun a reverse interview where the AI asks clarifying questions first, establish strict non-goals, store persistent project conventions in AGENTS.md, and provide version-locked documentation using Context7 and Repomix so the agent never hallucinates obsolete APIs.
Core Techniques:Reverse Interviewing (Superpowers)Persistent Project Memory (AGENTS.md)Version-Locked Context (Context7 & Repomix)- Key Deliverable
- A structured, one-page SPEC.md defining actors, data entities, failure states, and acceptance criteria.
- 01Stop Vague Prompting with a Reverse Interview1:45
- 02Give Your Agent Project Memory with AGENTS.md and Scoped Rules1:50
- 03Replace Outdated API Guesses with Context7 and Repomix1:48
02Engineer Your Agent: Skills, Context, and Worktrees3 lessons
Module 02 · Real-World ScenarioTreat your AI coding agent like an engineer who needs a clear handbook, an organized desk, and a safe sandbox.
You wouldn’t dump a 500-page operational manual into every short conversation with a teammate, nor would you let an intern run experimental tests directly on your production branch. When you connect too many tools or paste huge documents into an AI session, its memory context fills up, causing it to ignore earlier rules and make random mistakes.
⚠️ The Vibe Coding TrapPiling dozens of uninspected AI plugins and endless instructions into one massive prompt, hoping the agent will magically sort through the noise.
✓ The Engineering SolutionStructure instructions into three tiers: always-on project rules, on-demand Agent Skills (SKILL.md) that load only when needed, and isolated Git worktrees that give the AI a completely separate sandbox directory for experiments without risking your working code.
Core Techniques:Context Budget ManagementAuditing Vercel Agent SkillsGit Worktrees for AI SandboxingSemantic Symbol Search (Serena)- Key Deliverable
- A custom, auditable SKILL.md for database operations and an isolated experimental Git worktree.
- 01Manage the Context Budget with Rules, Skills, and MCP1:52
- 02Audit Vercel Agent Skills and Write Your Own SKILL.md1:44
- 03Run Safe Git Worktree Experiments and Find Symbols with Serena1:24
Phase 2Modules 03–04
Build the feature
- Outcome
- Publish a real incident update from an accessible form with verified data access.
03Design and Polish a High-Craft Interface3 lessons
Module 03 · Real-World ScenarioMove beyond generic "AI slop" by enforcing real design tokens, accessible motion, and surgical component targeting.
A stressed user opens your status page on a mobile phone under bright sunlight to see if their service is down. If the AI generated low-contrast grey text, erratic animations, and overlapping badges, the user cannot find the answer. When you tell the AI "make the badge look better", it often rewires the wrong component and breaks the entire screen.
⚠️ The Vibe Coding TrapAccepting generic, unstyled AI templates with jarring movement, missing focus outlines for keyboard users, and hard-coded inline colors.
✓ The Engineering SolutionDefine deliberate semantic design tokens (DESIGN.md) using tools like tweakcn and UI UX Pro Max, apply calm, purposeful motion with shadcn and Anti-Slop heuristics, and use React Grab in development to target the exact UI component node on screen.
Core Techniques:Semantic Design Tokens (tweakcn & DESIGN.md)Accessible UI & Natural Motion (shadcn & Anti-Slop)Visual Element Targeting (React Grab)- Key Deliverable
- A high-contrast, keyboard-accessible status interface that maintains clarity across mobile screens, light/dark themes, and reduced motion.
- 01Design a status interface with UI UX Pro Max and tweakcn1:40
- 02Add motion to an accessible incident form1:40
- 03Find the exact element with React Grab and check five UI states1:28
04Build Real Features: Data, Access, and Honest Feedback4 lessons
Module 04 · Real-World ScenarioNever trust a green checkmark or a loading spinner until you have verified the database write and security boundaries.
You click "Publish Update". A green toast notification pops up saying "Published successfully!" But when you open a private browser window, nothing has changed. The AI had simply written a fake timer (setTimeout) that showed a toast without actually persisting data to the database. Or worse: anyone on the internet could edit your company’s incidents because row-level security was missing.
⚠️ The Vibe Coding TrapMistaking a visual animation or happy-path UI simulation for a functional, secure backend implementation.
✓ The Engineering SolutionTrace the data pipeline from React form to backend server action to SQL database. Use Supabase MCP to inspect local schemas, validate migrations and regenerate TypeScript types, enforce role-based access control with synthetic accounts, and provide honest, recoverable error states.
Core Techniques:End-to-End Save TracingDatabase Migration Review (Supabase MCP)Row-Level Security & Role ChecksDraft Preservation & Honest Error Handling- Key Deliverable
- An authenticated publishing pipeline connected to real relational data, guarded by security policies and resilient to network dropouts.
- 01Follow a vertical slice from form to saved row1:22
- 02Review Supabase migrations and regenerate types1:45
- 03Verify organization permissions with Row-Level Security1:27
- 04Catch fake saves and surface real errors1:32
Phase 3Module 05 + final assessment
Verify and release
- Outcome
- Use browser evidence and code review to decide whether the release is ready.
05Verify and Ship: Browser Tests, Security Review, and the Launch Gate3 lessons
Module 05 · Real-World ScenarioThe AI will always tell you the code is ready. Only empirical verification and human judgment decide when to release.
It is Friday afternoon. The AI agent reports: "All changes implemented cleanly, ready for deployment!" But when a real visitor visits the live site, they receive stale cached information. Furthermore, a hidden security regression was introduced in an unreviewed helper function. Shipping without proof turns an AI shortcut into an outage.
⚠️ The Vibe Coding TrapTreating the AI coding agent’s confidence as proof of correctness and shipping without independent verification.
✓ The Engineering SolutionRun automated end-to-end browser tests using Playwright MCP and Chrome DevTools MCP to reproduce and verify behavior from a real visitor perspective. Run automated static security scans using Semgrep, conduct independent multi-agent reviews, and enforce an evidence-based release checklist with clear stop signals.
Core Techniques:Automated Browser Regression Testing (Playwright & Chrome DevTools MCP)Static Security Analysis (Semgrep & Semgrep Skills)Independent Code Review & Diff TriageEvidence-Based Release Decision Gates- Key Deliverable
- A repeatable browser regression test suite, a clean static security audit, and a documented release checklist with human sign-off.
- 01Use Playwright and Chrome DevTools to verify the browser1:36
- 02Review a diff in a fresh session with Semgrep1:39
- 03Make the release call with a green, yellow, and red checklist1:41
- Final release assessment12 scenarios
Final assessment and certificate
Work through 12 decisions about the service-status app. Review the agent’s changes, weigh the evidence, and decide whether to release, hold, or roll back.
Take the final assessmentPass all five module quizzes and the final assessment in the same browser to claim your signed SanX Academy Certificate of Completion.
Claim a certificateSanX Academy
SampleCertificate of Completion
Alex Rivera
Beyond Vibe Coding: Building High-Craft, Production Apps with AI
Coding agent toolkit
Find the worked lesson and official docs for each tool.
Superpowers
Ask the agent to interview you before it plans the first publish change.
Use the brainstorming or planning skill on ambiguous work. A short chat works for a small change.
AGENTS.md and Agent Skills
Put durable repository instructions in project guidance and recurring procedures in an on-demand skill.
Check your client’s discovery rules. Review bundled scripts before giving an agent permission to run them.
Context7 and Repomix
Check installed API versions against documentation and review a small public reference excerpt.
Context7 uses a hosted service with quotas. Review packed files before sharing them and inspect the exact installed version.
Vercel Agent Skills
Review React request dependencies and use the relevant skill to check the incident form.
Measure the actual request timeline. Apply guidance to the framework and code it fits.
Git worktrees and Serena
Separate an experiment from a branch under review and resolve a symbol to its declaration and callers.
A worktree shares Git data and does not isolate credentials. Use normal file search when it already finds the target.
UI UX Pro Max and tweakcn
Draft a design direction, export theme tokens, and check contrast and status labels in the running page.
Review exported values. The theme still needs checking in both color modes and on smaller screens.
shadcn/ui MCP, Motion Primitives, and Magic UI
Find the real Dialog in your registry and compare a restrained in-app entrance with an optional public heading reveal.
Inspect components already in the repo first. Check focus, dismissal, and recovery even when a library handles the primitive.
Emil Kowalski’s Design Skills and Anti-Slop
Review motion, shadows, contrast, and focus against the project’s DESIGN.md.
Treat the audit as findings to verify in the browser. Apply the app’s own brand and reduced-motion rules.
React Grab
Copy the selected element’s source context so the agent can find the correct status badge.
Use it in a development build. Check which parent owns the layout and review both screen widths.
Supabase CLI and MCP
Inspect a scoped development schema, review a migration, and regenerate types from the local database.
Keep sample data in a disposable environment. Hosted MCP uses project_ref for project scope, with separate read-only settings.
Playwright MCP and Chrome DevTools MCP
Replay the user journey, then inspect the actual network response to see where the stale update came from.
Use separate test profiles. Playwright WebKit gives compatibility evidence, with device-specific Safari behavior verified separately.
Semgrep Community Edition and skills
Review static findings alongside a fresh diff review and behavior checks.
Read the findings and coverage. Registry requests may use the network, and a successful command can still report findings.