This content originally appeared on DEV Community and was authored by Tim Wheeler
After 30 years of development, AI coding tools promised to revolutionize my workflow. After much learning, I found a solution.
In the Beginning
Do you find yourself restraining from scolding your AI developer? Well, if yes, you are not alone.
When I started working with Claude Desktop, then Claude Code, I said, “This is so cool!”.
Reality Hits Hard
“This is so cool!” quickly turned into “What a pile of @#$%#!”
The problems were everywhere:
- Inconsistent patterns across different sessions
- Hardcoded configuration scattered throughout the codebase
- Missing security validations that any junior dev would catch
- Architectural decisions that changed randomly between conversations
- Reinventing the wheel instead of using established libraries
Claude has epic amounts of Intelligence, and teenie tiny amounts of wisdom.
Giving Claude Wisdom
After much frustration, I had a realization: the fault wasn’t Claude’s—it was mine.
AI systems are basically a pattern matching behemoth. Without proper guidance to narrow down those patterns, they’ll match against everything they’ve ever learned. That’s a recipe for inconsistent, buggy, insecure code.
The solution? Persistent guidance, and the term in the industry right now is Context Engineering.
My Failed First Attempts
I started simple: markdown files with coding standards, architecture decisions, and examples stored in Git repositories.
This worked… sort of. But it quickly became cumbersome:
- AI couldn’t easily discover relevant guidance
- No way to ensure AI actually read the documentation
- No continuity in our conversations
- Guidance scattered across different repos and folders
- Lack of governance
- Compacted conversations lost a lot of the initial reading
I needed something better.
The Real Solution: Context Management
I built what I call the Universal Context Manager (UCM) – essentially “GitHub for AI-consumable artifacts.”
Here’s what it does:
- Structured storage for guidance, templates, and examples
- AI-friendly tools through MCP (Model Context Protocol) servers
- Discovery-first workflow so AI finds relevant patterns automatically
- Version control for guidance and templates
When I first started using the UCM for my own work, it felt like I was 7 years old again and got my first bike. Wow, this is really cool. Let’s make it better.
The aim was to minimize friction between the AI and the system.
So, after much trial and error, working with Claude, testing, reviewing and improving the workflow, the UCM was ready.
Instead of inconsistent code, I got:
- Quickly get up to speed with the ‘Recent Artifacts Tool’
- Find code, patterns and guidance through the ‘Search’ and ‘List Artifacts’ tool
- Consistent architectural patterns
- Proper configuration management
- Security validations by default
- Reusable, tested components
What was so powerful about the UCM:
The MCP Server tools forced Claude to do a few key things:
- Get initial guidance from the ‘UCM Quickstart’ tool which explained important concepts and rules
- Forced Claude to assign Categories and Subcategories to the artifacts, creating a natural structure and coherent filing system
- Gave Claude valuable metadata to fill out, such as a ‘Description’, allowing for search and quick retrieval of relevant details.
And all this I get now for almost no effort. Claude just knows how to deal with things.
This is the closest to ‘J.A.R.V.I.S.’ I’ve been able to get, one of my goals 🙂
A Real Example: The MCP Server Saga
Let me share a concrete example that shows the before and after.
The Task: Build a Remote MCP Server compatible with Anthropic’s specifications.
Without UCM (Days of pain):
Day 1: Claude builds NodeJS server → Incompatible with NextJS
Day 2: Rebuild for NextJS → Doesn't follow MCP protocol standards
Day 3: Add Anthropic MCP SDK → SDK incompatible with NextJS
Day 4: Build custom adapter → Partial success, still buggy
Day 5: Discover 'mcp-handler' package → Finally works!
Many days of trial and error, wrong turns, and… let’s say “colorful” feedback to Claude.
Next time, With my published guidance in the UCM (30 minutes):
Me: "Hi Claude, build a Remote MCP Server. Check UCM for implementation guidance."
Claude: *Reads stored guidance, finds mcp-handler solution, generates working code*
Result: Working server in first attempt
Check it out here: nextjs-remote-mcp-server.md and here route.ts
Claude remembers your history
I got a little frustrated with having to repeat myself, you know even a Junior Dev get’s it after some practice, but not Claude. That’s solved now.
The Game Changer: Shared Knowledge
Here’s the powerful part – I documented that entire MCP server solution in UCM’s public repository.
This means other developers can:
- Get a free UCM account
-
Add the Remote MCP Connector:
https://ucm.utaba.ai/api/mcp - Ask their AI to check the Utaba repository for MCP guidance
- Build a working Remote MCP Server immediately
No more three-day debugging sessions. No more reinventing solutions.
Building a Community
The vision extends beyond personal productivity. Imagine a community where developers share battle-tested guidance:
- Security patterns that prevent common vulnerabilities
- Architecture decisions that scale properly
- Integration solutions for complex scenarios
- Best practices that actually work in production
Currently, UCM accounts are private (you access your guidance plus public Utaba patterns). But I’m exploring public repositories – carefully, because we all know the security implications.
What’s Next?
I’m looking for developers who want to try this approach. There’s a free tier that will always exist, plus I’m offering free paid subscriptions for anyone willing to test the system and provide feedback.
The goal isn’t to replace your development skills – it’s to amplify them. When your AI has consistent, tested guidance, you spend time solving business problems instead of debugging inconsistent generated code.
More importantly, you get a code base built on standards, that is cohesive and consistent.
Connect to the UCM and become a true orchestrator of AI Development.
Want to try it?
- Grab a free UCM account
Add the Remote MCP Connector:
https://ucm.utaba.ai/api/mcpJoin our Discord for feedback and discussion or to request a free upgrade subscription: https://discord.gg/ZkVEJvBt
Send feedback directly through your AI (yes, really – we built tools for that)
After 30 years of development, I thought I’d seen every productivity tool. Turns out, the missing piece wasn’t better AI – it was better context management.
By Tim Wheeler, Code Monkey, Solution Architect – Passionate about Tech, Founder of Utaba.
Tags: #ai #claude #development #aigovernance #productivity #mcp #coding #contextengineering #anthropic
This content originally appeared on DEV Community and was authored by Tim Wheeler