Introducing AI Skills for Real Engineers
Introducing AI Skills for Real Engineers êŽë š
AI is an undeniable force in software engineering right now. These tools are trivial to use: tell it what to do, and it does it. But as easy as these tools are to start using, they can be tricky to use well.
This is a post about âAI Skills for Real Engineersâ (mattpocock/skills) which is a suite of AI âskillsâ (the kind you install into your AI harness) created by Matt Pocock. These skills exist to help you make AI tooling more effective. Iâve found them to be straightforward to use and very impactful.
Weâll take a brief tour through a few of these skills, which I think most clearly shows the kind of value they can add to almost any software devâs workflow, no matter what youâre working on.
Installation
Iâll assume youâre using Claude Code (see the docs for other harnesses). To install:
claude plugins install mattpocock-skills
If that fails with something likeâŠ
"mattpocock-skills" not found in any marketplace
Then try clearing your Claude Code Marketplace cache (yes, really). This command will do just that:
claude plugins marketplace update
A Skill to Help You Use the Skills
The docs for these skills are pretty clear, but you might still have some questions. Believe it or not, thereâs a skill that helps you understand what the other skills do.
Letâs check it out. Imagine you read the docs and youâre not quite sure what the difference between the /grilling and /grill-me skills is. You can just fire up the /ask-matt skill, and ask it.
Letâs actually put these skills to work in an actual project.
Project Setup
Many of these skills will do things like create tickets, or even documentation like ADRs. To get those, and other things set up, the first thing you should do in a new project is run /setup-matt-pocock-skills.
This is where we configure these skills: where to create issues, where to create ADRs, and so on. Itâs a simple thing, but itâs a nice touch to help the other skills run more smoothly.
Producing Clear Requirements with /grill-me
Anyone whoâs used LLMs for coding knows that clear, detailed prompts are essential. Missing details are anathema to effective AI use. If you leave AI to assume things youâve left out of your prompt, you might be disappointed in the result. Tools like Claude Code do have a plan mode, and LLMs in general will happily accept things like âDid I miss anything?â at the end of your prompt, but thereâs a better way.
The /grill-me spec formalizes all that and takes it to the next level.
To get started, just do /grill-me and describe your feature.
Itâll analyze your prompt and come up with some surprisingly detailed questions. As you answer those, youâll likely be greeted with some follow-ups.
Itâll keep going like that until it has what it needs.
Implementing (Now, or Later)
At this point, your session and context should have everything needed to implement your feature. You can absolutely feel free to tell Claude something like âlooks good, build it.â
Or if, for whatever reason, youâre not ready for this feature to be implemented right this secondâperhaps you have 2 or 3 other AI-generated PRs to test and review, perhaps you have two other agents building things right this second, and were just using that waiting time to spec the next thingâthen read on.
Saving Work for Later with /to-spec
If youâd like to take the entirety of the current conversation and context and turn it into a single issue for later, you can use the /to-spec skill. Just call it up, and let the skill do the rest. Itâll even try to add some tests and check with you about the appropriate testing boundaries.
Saving Work for Later with /to-tickets
What if the feature you just designed is big? Humans work best with small, well-defined tasks, and AI agents are no different. Youâll likely get better results if you avoid letting your context window get flooded with content you wouldnât otherwise need.
Inside that same conversation you just had, via the /grill-me skill, you can call up the /to-tickets skill, which will break that feature into multiple issues for you.
Itâll even be smart enough to block tickets as needed, based on dependencies. Naturally, you can make any tweaks to the proposed result youâd like.
Once youâre happy, tell it so, and itâll do its thing.
Youâll wind up with a nicely filled-out board.
Learning with /teach
Weâve all used LLMs to help us learn or understand something. This skill suite actually has a skill that takes it to the next level. Fire up the /teach skill, tell it what youâd like to learn about, and itâll actually put an entire lesson together for you.
When itâs ready, the lesson will pop up in your browser.
When youâre done with that, it can keep churning on the next lesson in this topic.
Parting Thoughts
I hope you find these AI skills as useful as I do. They can really help refine and clarify your ideas into clear, actionable specs that your LLM can execute most effectively. This post has barely scratched the surface of everything they can do. Check the docs for a fuller picture!