One of the requirements of my job is to learn to use the agentic coding tools (Claude, and GitHub Copilot, specifically), and over the last three weeks, I feel like I've made huge strides in my understanding of the tools, their limitations, and where there's some cool stuff that can be done.

This weekend, I've been working on some personal projects and having a good time with my "unlimited" benefit that my work is providing me, and has expressly said "go use for your personal projects, too." One of the things that has come out of this weekend is that I'm finding some lessons that will be useful in my work day, which of course, is why they want people to take the stuff home with them.

First up is that I spent Saturday playing with Claude on a personal API-based subscription, and I installed a bunch of plugins using the built-in marketplaces. There are a lot of choices, and I selected what looked like a good set of picks for building a static website hosted in Azure. Claude has a much better marketplace/plugin/mcp navigation UX than Copilot. What I didn't realize, though, was all the design and planning I was doing (slowly reading each command, and output) ended up just being "theory crafting" using some of the skills. It didn't actually output any code into files. Oops! It did help me develop a solid plan, figure out some testing strategies, etc.

The second thing I have been learning/digging deep into, is the difference between "user" and "project" instructions and configurations. I think there is a general set of plugins/skills/agents/mcp servers that I want to use for a specific project, but what I've been doing at work is implementing them all in the user context, and this has been making my environment really unwieldy. I don't really need 14+ browser prompts to log into every MCP server every time I launch Copilot. I think what I'm going to do is similar to this mornings exercise, where I installed the "awesome-copilot" plugin from the "awesome-copilot" marketplace that's built in to Copilot. This provided a set of skills to help parse the marketplace and find skills relevant to the project I am working on. This was exceptionally helpful when I cloned my project and said, "based on these project plans, let's get a list of helpful skills, agents, plugins, and MCP servers." This was fantastic at helping me navigate and parse the awesome-copilot repo (not the least because the navigation of the plugins leaves a lot to be desired).

Lastly, I can't stress enough the need to iteratively re-evaluate any plans, designs, and architecture multiple times, with different agents, and refining the skills and agents that are being used as part of the review. Each cycle, I was able to develop closer to the kinds of standards I use in my work-day development, without having to export the tribal knowledge into some kind of KB. At the end of the current build process, I'm going to out put the development environment into a package that I can share with coworkers, who can use this as a starting point for their own exploration and journey.