The Podcast Page

After last week’s struggle with local vibe coding, I shifted my focus to remote models. The result? You can now see a live podcast page on this site. Here is how it went down.

From “Hard Work” to “Skills”

As an avid podcast listener, I often struggle with episodes in Russian or German. My goal was simple: send a Pocket Casts link to an LLM, have it find the episode on Apple Podcasts (which provides transcriptions), and generate a summary.

Initially, this was a heavy lift for OpenClaw. Running the sequence through Claude 4.6 Opus cost roughly $0.30 per summary. However, we are now in the “Skill Era.” I instructed OpenClaw to save those logic steps into a reusable Skill. Now, the heavy lifting is handled by Gemini 2.5 Flash Lite. It’s incredibly efficient; even after multiple runs, the cost hasn’t even hit half a cent.

The “Unsupervised” Myth

Then came the true “vibe coding” test. I used Claude 4.5 Haiku with OpenCode, attempting to stay hands-off. You’ve likely seen videos of developers “unsupervised” while agents build entire projects. Honestly? It’s fake.

The biggest hurdle is the Agentic Loop. As a project grows, the context expands, and agents often lose their place and repeat tasks. Without supervision, an agent will happily burn through your credits while running in circles. My “hands-off” attempt cost 30 cents and didn’t even deliver a basic player widget.

My dream was to set a plan at night and wake up to a finished product. We aren’t there yet.

The Good, The Bad, and The Messy

When I finally stepped in to supervise the creation of the podcast card you see on this blog, the code quality was… questionable.

• The Ugly: The agent used an excessive amount of any types and redundant if-else blocks. This bloats the context window—which, cynically, earns the LLM providers more money via token usage.
• The Good (Skills): Summarizing prompts into “Skills” is a game-changer for repetitive tasks. It keeps the context clean and the costs down.
• The TUI Comeback: Because LLMs are such resource hogs, running a heavy IDE alongside them is a struggle. Switching to LunarVim (TUI) was buttery smooth. As long as memory shortages exist, Terminal User Interfaces will remain king.
• TDD (Test Driven Development): I’m using LLMs to generate unit tests as snapshots.

The “vibe” is getting stronger, but for now, the developer still needs to keep one hand on the wheel.