I've been using Claude Code for two weeks now, and I haven't written a template yet.
That surprised me initially. I thought I'd need to formalize my prompting approach. Create reusable templates. Build a library of prompt patterns. The way you'd build a library of code snippets.
But that's not what's happening.
Instead, I'm doing something much more chaotic and somehow more effective: I'm going prompt to prompt, project to project, adjusting based on what I'm trying to accomplish. It's less about templating and more about learning the language of asking Claude Code good questions.
The typical flow is: prompt, let it run for 20-30 minutes, come back to the laptop and look at what it produced. Then either: "This is great, ship it," or "This is close but here's what needs to change," and iterate from there.
No templates. No validation frameworks. No meta-layer of tooling.
Just: good intent, clear context, and letting the AI work.
What I'm realizing is that the bottleneck isn't the mechanics of how I communicate with Claude Code. The bottleneck is whether I'm clear about what I actually want.
When I write a vague prompt, I get vague results. When I write a prompt where I've clearly thought through the problem, I get good results. The tool amplifies clarity. It doesn't create it.
That's different from code generation tools where there's often a mismatch between what the tool is optimized for and what you're trying to build. With Claude Code, the mismatch is usually between me and myself — between the clarity I think I have and the clarity I actually communicate.
I could build a framework to solve that. Templates and validation steps and structured prompts. And maybe eventually I will. But right now, the friction of that framework would probably exceed the value it provides.
Instead, I'm building something less visible: intuition. Pattern recognition about what prompts work. Understanding of what Claude Code is good at and where it struggles. The feel for when I need to be more specific and when I can be more abstract.
That's the tacit knowledge I mentioned in an earlier post — the stuff that can't be templated. The stuff you only develop by doing the work repeatedly and paying attention to what works.
One observation though: I am getting more consistent results. Not because I have a template, but because I'm getting better at the intent-setting part. I'm learning to think like the AI thinks. To decompose problems the way Claude Code will decompose them. To anticipate where it will get stuck and pre-emptively clarify those points.
That's a skill, not a system.
And skills are harder to teach and harder to scale, but they're also more flexible and more powerful than systems.
So for now, no templates. Just me and the AI, learning to work together.
When that breaks down, I'll formalize. But I don't think it's going to break down anytime soon.