Ideas for a "Vibe Coding" Tool

Date: 29 Dec 2025 Type: Cleaned Transcript

Description

The user outlines a project to build slash commands that help users recognize when they are stuck in unproductive coding patterns. The key idea is to recognize when the AI tool is not progressing and to reframe the problem.

---

Transcript

Idea and Challenge

I'm just capturing some quick context data for this particular project, which I think will just be a couple of slash commands. But while it's fresh in mind, I thought it would be a good time to just capture the idea and the challenge, which I'm sure anyone who's spent a lot of time this year "vibe coding" is familiar with.

I would call it tunnel vision. It's when, to give the example that spurred this, I'm trying, I'm getting a backend for my AI podcast generation utility deployed, and we're trying to debug authentication on the side of CloudFlare and Google Cloud Run.

So I'm working, of course, with my sidekick, Opus 4.5, and we're trying new different headers, injecting them into CloudFlare for authentication. I was rolling with it for a little bit, and then I'm like, wait a second, this is way too complicated. Let's take a step back. We're just trying to get authentication set up for this app.

Now, what I've learned from doing a lot of vibe coding over the past year is that these pivot points are critical. The mistake that I've made so many times with these tools is, because AI tools are designed to be helpful and confident, they kind of lead you to think, yeah, yeah, we're just on the brink of fixing this. And it's when you as a human operator say, hang on, this is like we're going round—it's not even going around in circles—it's that this is not a constructive direction. We need to rethink.

The Challenge

Now, the actual mechanism that I'm looking at for doing that is, this is where the challenge is. I've seen some MCPs for this. In the current case, I'm looking at the context load here. It's about 31%. No, 31% left. So we're at about 70% context use with Claude, and at this point, I think above, I feel above 50, it starts to get kind of hit and miss. I would say at this point my, what I would gravitate towards doing, is saying, and this is where I think it's probably an agent tool. This is what we've tried. This didn't work. What's a better way?

Now, if I do that at this point in the conversation, where we've got, we're working with a cluttered context window at 70%, that's not where the model is going to do its best reasoning. So frequently my unstucking hack is to start a new thread saying, we did this. So in other words, I'm kind of doing human summarization. I'm saying, this is the problem that we're working on. This is what we tried, this failed.

You know, we were overthinking it, or you were overthinking it a bunch. Let's just think, think from, think of a new idea and then let's try to get that working. So the slash command might be that. Just carefully worded slash command intended for the next conversation, saying, okay, we were working on this. You know, and it might be an agent that I give this quick summary to, like, you know, maybe new eyes or new lens or something. And then it summarizes that and kind of hands it over as like a brief for the next agent, the next agent, of course, being just Claude Code. But that's basically the idea that I have for this. And what I've seen again with Claude so much, and all of these, is that really struggling for tooling actually in this regard. It's just that context management is the key. It's it's knowing there's no built-in tooling at this point. That tells you, okay, you're, we're getting, things are like we're not getting closer. We're this isn't productive. We're just, the context window is growing and we're trying more and more stupid things. And we need to start a new thread and to give it determinative guidance. And that's the specific tool or slash command or agent that I'd like to have at my disposal and quick reach for knowing at this point in the conversation, that can actually be a sanity saver because, minus this, I know from experience it will happily spend another hour trying out the same flawed idea of, let's try this different header in CloudFlare. And the thing is it'll probably get, it'll probably work, but using PhD level reasoning or higher reasoning to get an illogical work around to work is not actually success. It means we've, it's like we've built a, you know, we've cooked a pizza by ordering, instead of ordering from a local pizza shop, we've hired a chef by the hour and taking an industrial quantity of tomatoes. Stupid example, but that's the point. It's, it's, it means that we've achieved success but through a ridiculous mechanism, and that's not really success. So that's the, these are the placeholder notes for working on this slash command.