Fuzzwah's Agentic Coding Learnings

Three weeks ago I found myself staring at three different “Product Requirements Documents” that were supposed to guide my development work. Each one written in a different style, with different levels of detail, and — most frustratingly — each left critical questions unanswered.

The first was essentially a wall of implementation details. It told me how to build the feature (database schemas, caching strategies, partitioning approaches) but buried the why so deep I had three different interpretations of what problem I was actually solving. The second was the opposite — eight pages of ambitious vision with phrases like “engaging experience”...
Continue reading →

I’ve been messing around with AI-assisted development for months now, and after a frankly embarrassing amount of tinkering I’ve landed on a setup that actually works. I can already feel your eyes preparing to glaze over, but stick with me — this stuff is genuinely exciting and has completely changed how I build things.

The Foundation: VS Code Insiders

I run VS Code Insiders rather than the stable release because:

• MCP support and Copilot improvements land here first
• It’s surprisingly stable for a daily build
• I can keep stable VS Code installed side-by-side for critical work if...
  Continue reading →

I’ve spent way too long messing around with MCP servers at this point, and one of the first decisions you hit when building or choosing one is the transport: STDIO or HTTP. Sounds like a boring infrastructure detail, right? Turns out it fundamentally changes how people actually use your server. Let me share what I’ve learnt building and deploying both.

STDIO Servers: The Automatic Advantage

The killer feature of STDIO MCP servers is dead simple: VS Code automatically starts them for you.

When a user adds your server to their settings.json:

So I gave an AI agent unrestricted access to my database and it went ahead and restructured my schema. Surprise, surprise. Let me tell you how that little disaster led me to build a whole new access mode for postgres-mcp.

The Problem: Two Modes Aren’t Enough

The postgres-mcp server shipped with two access modes: UNRESTRICTED (full database access, great for dev, terrifying for anything else) and RESTRICTED (read-only, safe but useless when you actually need to write data). What was missing was the middle ground — let the agent insert records, update statuses, delete old data, but for...
Continue reading →