Restarting Blogging 3rd Time , What's New This Time

Think about what a how-to post was actually for. You had a problem, you searched, you found someone who'd hit it before, you adapted their answer to your situation. Every step of that is now worse than asking a model. The model answers the version you're running rather than the version I was running in 2020. It answers against your code rather than my sample project. It doesn't make you skip four paragraphs of preamble.

That's the same staleness problem from a different angle. My posts aged because they were pinned to a version.

So the genre is commoditised, and I don't think that's a loss worth mourning. What it does mean is that publishing now requires having something a model can't produce. A model can tell you how the API works. It can't tell you what broke when I ran my own framework on a different harness, or what a decision cost me three weeks later, or which approach I tried first and abandoned. That's first-hand, and it's most of what's left.

What this is instead

I'm not going to promise a cadence. I've made that promise before and I don't have evidence I can keep it. This is an experiment, and I'd rather you judge it on what shows up than on what I said I'd do.

What I can describe is the kind of thing that shows up. Every post here is the writeup of something I actually built. Not a topic picked for an audience, but the work itself, documented after the fact:

• Something I built, with the design decision and the alternatives I rejected.
• Something that broke, with the diagnosis and the mistake left in.
• Something it cost, with real token and time figures from real runs, reported as ratios.

Two reasons for that shape. It's the part a model can't write, since it depends on having actually done the thing. And it makes the writing a by-product of work I'm doing anyway rather than a second job, which is the only version of this I have any reason to think I can sustain. If I haven't built anything worth writing up, there's no post, and that's a truthful signal rather than a gap to paper over.

What's already in the queue

I know what the next post is, because it already happened. I've spent a good part of this year building an agentic development framework. When I ran it under a different agent harness, it broke, badly enough that calling it a framework was premature. The writeup is what was coupled to what, and how I missed it.

There's a symmetry there I'd rather not have spotted. The archive was coupled to a tool. The framework was coupled to a harness. Twenty years apart, same mistake. The only improvement is that I caught the second one before publishing the claim instead of after.