How I Actually Use Claude, Copilot, and Grok (today)
I've used AI in my development workflow every day now for about two full years. Most of that work at this moment happens across three tools: Copilot, Claude, and Grok. Developing your prompting skills is the logical first step: it's your interface to the model, and better prompts get you better results. But to really benefit, you need to build skills: reusable prompts that you can run again and again.
I started building skills in Claude for the things I do constantly: reviewing code, generating tests, and a handful of stack-specific helpers for the stacks I work in. Grok I keep open for a second opinion when Claude and I are talking past each other. No, I don't use OpenAI currently, although I have in the past and may again in the future. There's only so many models I can pay for!
The repeated prompt is the signal
The first time I asked an AI to review a diff, it was a throwaway prompt. By the fourth time I typed some version of it, I caught myself hand-rolling the same prompt over and over. So I wrote it down properly: a code-review skill that actually describes what I care about in our codebase. Now I run it instead of retyping it.
Tests went the same way. I have a skill that knows how we write tests on this stack (JUnit 5), what we mock and what we don't bother mocking. There's nothing clever about it. It's a prompt I've iterated on a bunch until it stopped getting things wrong.
Narrow beats ambitious
The skills that flopped were the broadest ones. "Be my coding assistant"--not useful. The skills that stuck are more narrow and opinionated: "review this Java 21 code, ensuring it follows our conventions and is using the latest language features"; "scaffold a JSF page the way we actually build them".
Narrow means I can trust the output without rereading every line, and it means I can fix the skill when it drifts.
Smaller skills cost less
There's a cost angle people miss. Every call ships the skill's instructions along with it, and you pay for those tokens. A do-everything skill drags a wall of context into every request whether that request needs it or not. Break it into a few focused skills and each call carries only what's relevant, so you spend fewer tokens.
It adds up faster than you'd expect when you're running these all day. Tighter skills also tend to land the right answer on the first try, which spares you another few rounds of tokens spent talking the model back from a bad start.
You still own what it writes
The thing that bites people is overconfidence. AI will happily give you a wrong answer with the same certainty as a right one, just as it will go along with a bad idea if you propose it confidently enough. That agreeableness is a real hazard when you're the one deciding what ships.
So I push back in writing. Every project gets a CLAUDE.md or AGENTS.md, and somewhere in it I tell the model to challenge me when I'm wrong and call out the holes in my plan instead of just agreeing.
State assumptions. Surface tradeoffs. Ask before guessing. Push back when a simpler approach exists.
The speed is real, and everybody already knows that. What people skip over is that the accountability didn't move an inch. If it ships broken, that's on me.
They need upkeep
A skill I wrote six months ago might not hold up today. The models change, our codebase moves, and a skill that was sharp goes dull. Every few weeks I go back through them, delete the ones I've stopped using, and rewrite the instructions on the ones I lean on. It's maintenance, same as any other code I'm responsible for.
AI can help you maintain skills
One of my favorite things to do is ask the model to review my skills. I'll paste in the instructions for a skill and ask if it still makes sense, if it's missing anything, or if it could be clearer. Then, I have AI interactively ask me questions so we can iterate on it together. This helps me keep the skills sharp without having to do a full rewrite. The key here is reviewing the changes as you go along to make sure it still makes sense. This is where being concise with your skills really helps...if a skill gets too long, it may be time to split it out into smaller, more focused skills.
That's the part I'd pass on to anyone getting started. Stop hunting for the perfect prompt. Build a few small skills for the work you repeat, keep them in good shape, and read everything they hand you.
