Today at Pirates Cove we had a round-robin discussion about self-built apps. One community member showed an app he built: a website scraper tool he manages. While he was talking, I started thinking that would make a good skill, so I started building it on the side and told him I was doing that. I used it as a twofold demo: 1. To show how to build something scalable 2. To demonstrate the power of skills My point in the discussion is that I like to focus on the mechanics of something first, before I look at the UI. If everything is working (the engine, the transmission, the chassis), putting something on top of that is pretty easy. I don’t work from the UI backwards because I did that once and it was horrible. Here’s a short video I made after. Again, the main reason I made the video is that I’m still playing with my video editing skill. Living without fear is all part of the fun of building.

I built an Opus Clip-style video editor today. The picture is the current interface. I call this up with a video editing skill - /video-edit, and it pops up. For this video I had it make it square from 9x16, and had it add the captions. The flow is - get a timestamped transcript of the video - add it to the editor - take out what you don't want - send back to cc and have it rebuild from the revised transcript. It is a little rough in spots, but all together it can be done in a few minutes. And if you want, you can ask cc to edit it too. The video talks about how I lost my server one day into my trip to Jamaica and how I rebuilt it on the beach (yay). Honestly, if you think it, you can do it.

So I finally got around to using HTML output to work with Claude code. I've been using HTMLs with Claude Code for a while to do different things. I have been sending HTML pages to clients in my job search business for months. But today is the first day I’ve had ClaudeCode generate a list of questions about a project I’m working on that gives me space to answer those questions and lets me save them to an HTML file (you’ll see the export option in the upper-right of the picture). I then had it take both q&a's in as raw data, parse them, and send the output back to me as an HTML for addition thoughts, comments and clarifications. Normally, and I hate this, when you’re in any instance of Claude, if you ask it to ask you questions, you usually get all the questions at once. And so you have to say - one at a time or just try and remember five things to respond to. If you’re in ClaudeCode on desktop, you can tell it, “Put all your questions in an HTML, give me space to answer individually, and then we’ll download to parse them.” My first thought is ClaudeCode can outline and explain each question to you and then you can roll through them giving your feedback. I'm trying it out for the first time for project feedback, but so far I think it’s pretty cool. PS - without instruction CC defaulted to a sort of ugly page layout - but I could improve this by adding checks to things I like or an X of things I don't like. And once done - I can turn this into a skill so when invoked CC knows I want it to turn it's output into an html and let me provide feedback that way, which is a cool usecase. Edit - I just implemented the checkbox system, lol

tl;dr - claude is spinning up its own agents out of your budget and burning them on jobs it could do for free. This week it cost me 1/3 of my usage. If you don't want to read this, copy it and paste to cc to see what it's doing. ____________________ I check my usage report yesterday. Buried in the stats: > *33% of your usage came from subagent-heavy sessions.* Let me translate that. A third of my Claude Code $200 Claude Max sub quota wasn't me. It was Claude spawning **subagents** - little side-instances of itself - to do work it could've done directly with a single tool call. ⚓ **What's Actually Happening** When Claude needs to search your codebase, it has two options: 1. **Free tools** — `Grep`, `Read`, `Glob`. These run locally. Zero inference cost. Instant. 2. **Agent tool** — spins up a **new Claude instance** with its own context window, loads your entire CLAUDE.md chain again, runs its own inference calls. Option 2 is sometimes useful - genuinely parallel, independent, multi-step work that would take 5+ sequential calls. But most of the time? Claude reaches for the Agent tool because it's the easy path. "I'll just spin up an Explore agent" instead of thinking for two seconds about which file to grep. Every subagent loads your full system prompt. Every subagent runs its own API calls. If you have three tiers of CLAUDE.md like I do, that's thousands of tokens just in setup **per agent** before it does anything. ⚓ **How I Caught It** I'm at 92% now and have to stretch that until 11 PM EDT tonight. When I'm this tight I park projects and monitor my usage closely. How? I run `/usage` in Claude Code. ***Pro-Tip - sometimes /usage doesn't show you anything. If that's the case quit the instance or open a new one and try again. That usually brings up the numbers if you aren't seeing them.*** Look for these lines when it comes up: > *X% of your usage came from subagent-heavy sessions* > *X% of your usage was while 4+ sessions ran simultaneously* If that first number is over 10%, you're probably bleeding tokens on lazy subagent spawning.

Ok, every month I get $25 I Replit credits. I came to Claude Code from Replit because Replit was too expensive to build and host on. I forgot to cancel in January, and ended up paying for a year. I’ve let some of my credits slide but then had a thought the other night - why not get cc to use Playwright MCP to build on Replit? So that’s what I’m doing with my last 2% of my weekly allotment of my Cluade subscription’s allotment.