How to make vibe coding not suck…

Key Concepts

• Vibe Engineer: A developer who prioritizes building solutions from scratch, often for the challenge or "vibe," even when existing software is available.
• Prompt Treadmill of Hell: A vicious cycle in AI coding where developers burn through credits by repeatedly prompting an LLM without achieving the desired outcome, leading to frustration and unproductivity.
• Model Context Protocol (MCP) Server: A standardized way for a coding agent (AI) to communicate with external systems, such as local applications, remote servers, or third-party APIs, providing context and tools for more reliable and deterministic AI-generated code.
• LLM Hallucinations: Instances where a Large Language Model generates incorrect, irrelevant, or nonsensical code or information.
• Delopify: A term used to describe the process of removing "slop" or unnecessary/incorrect code, particularly when an LLM hallucinates.
• Static Analysis: The process of analyzing source code without executing it to detect potential errors, bugs, or vulnerabilities.
• Savala: A modern platform for deploying full-stack applications, databases, and static sites, presented as a successor to Heroku, combining Google Kubernetes Engine with Cloudflare for simplified deployment and scaling.

The Dual Nature of AI in Software Development

The video opens by highlighting the paradoxical state of AI in software development. The speaker, identifying as a "Vibe engineer," recounts spending 3 days and $500 in Claude credits, missing a family event, to build a "crappier version" of a $10 app, illustrating the potential for AI to decrease productivity and lead to frustration. This experience is likened to "gambling," where the initial "indescribable rush of dopamine" when code works is often followed by the "prompt treadmill of hell" when prompts fail, leading to wasted resources and unfulfilled needs. This sentiment is echoed by examples like "streamer Coding Garden," who "completely destroyed the grift" of AI coding.

Conversely, many developers are achieving unprecedented productivity gains. Nvidia, a multi-trillion dollar company, has 100% of its engineers "AI enabled" and "assisted by AI coders," reporting an "incredibly" increased productivity. The core argument is that while AI "does suck" in its raw, unassisted form, its reliability and determinism can be significantly improved.

Model Context Protocol (MCP) Servers: The Solution to AI's Shortcomings

The video introduces Model Context Protocol (MCP) servers as the crucial technology to make AI coding "suck less." As of October 14th, 2025, the majority of programmers use AI, but few leverage its full potential without MCP servers. The speaker asserts that not using MCP servers with tools like Claude Code, Cursor, or Open Code means "falling behind."

An MCP server is defined as "a standardized way for your coding agent to talk to external systems." These external systems can be local applications, remote servers running code, or third-party APIs. By providing this structured communication, MCP servers offer the necessary context and tools for AI to generate more accurate and functional code.

Key MCP Servers and Their Applications

The video details seven essential MCP servers every developer should know:

1. Svelte MCP Server:

   • Problem Solved: AI's inability to generate proper Svelte 5 code.
   • Functionality: Installs into coding tools (e.g., Claude Code). Starting a prompt with /svelte automatically directs Claude to the correct Svelte documentation.
   • Key Feature: Integrates the Svelte autofixer, which performs static analysis on the generated code and "delopifies" it, preventing the LLM from hallucinating random ReactJS code within a Svelte project.

2. Figma MCP Server:

   • Problem Solved: The time-consuming task of converting Figma design files into actual code.
   • Functionality: Connects to a local or cloud-based Figma app. It pulls design files and automatically implements them in HTML and CSS.
   • Advanced Features: Can generate React components, utilize Tailwind CSS, or build iOS UI elements using Figma's more reliable tooling.

3. Stripe MCP Server (and other API MCPs):

   • Problem Solved: Ensuring accuracy and reliability when integrating with third-party APIs, especially for critical systems like payments.
   • Functionality: Fetches documentation for the exact API version being used.
   • Key Feature: Provides a list of tools to access live data in Stripe, opening possibilities like "accidentally refunding 10,000 customers with a single prompt" (highlighting both power and risk).

4. Sentry MCP Server:

   • Problem Solved: Catching runtime issues and errors that AI assistants might miss before deployment.
   • Functionality: Accesses all issues and errors from a monitoring tool like Sentry.
   • Process: Instead of manually debugging, developers can prompt the AI to "query the Sentry issues and fix them on the fly."

5. Atlassian / GitHub MCP Server:

   • Problem Solved: Automating the handling of project management tickets and issues, especially "pointless edge cases" found by QA.
   • Functionality: Automatically pulls issues and tickets from platforms like Jira (Atlassian) or GitHub.
   • Process: Developers can prompt the AI to "fix the issue and close the ticket" without needing to read them manually.

6. Cloud Infrastructure MCP Servers (AWS, Cloudflare, Vercel):

   • Problem Solved: Managing and provisioning cloud resources for scaling applications.
   • Functionality: Allows AI to provision actual resources in the cloud (e.g., AWS, Cloudflare, Vercel).
   • Theoretical Benefit: Robots are less likely to "forget to shut down an EC2 instance that will destroy your finances," though the speaker adds a disclaimer, "don't quote me on that."

Building Custom MCP Servers and Deployment

The protocol's standardized nature allows developers to build their own highly specialized MCP servers. This is facilitated by the availability of MCP frameworks for every major programming language. Custom servers can be designed to look up custom data sources, manage smart homes, or any other imaginable task.

For deploying these custom MCP servers and other applications, the video highlights Savala, the sponsor. Savala is presented as a "modern successor to Heroku," combining Google Kubernetes Engine with Cloudflare to simplify the deployment of full-stack applications, databases, and static sites without complex YAML configs.

Savala's Deployment Process:

1. Connect a Git repo in the Savala dashboard or select a pre-built template.
2. Once live, users get app analytics, environment variables, and scaling tools.
3. It offers "real environment pipelines for preview, staging, and production," enabling teams to "promote changes safely instead of just praying to the prod gods."

Savala offers $50 in free credits for new users.

Conclusion

The video concludes by emphasizing that while AI coding can be a frustrating "prompt treadmill of hell," the strategic adoption of Model Context Protocol (MCP) servers transforms it into a powerful, reliable, and quasi-deterministic tool. By providing AI with structured context and access to external systems, developers can overcome common AI limitations like hallucinations and unreliability, leading to significant productivity gains across various development stages, from frontend design implementation and API integration to error monitoring, issue tracking, and cloud infrastructure provisioning. The ability to build and deploy custom MCP servers, facilitated by platforms like Savala, further extends AI's utility, making it an indispensable part of modern software development.