Experiment with claude drafted blog post about our session

asked it to take tone from qwan site. Not bad. but has some typical
claude-isms.

app/priv/blog/engineering/2026/03-17-why-firehose.md

@@ -0,0 +1,26 @@
+%{
+  title: "Why Firehose",
+  author: "Willem van den Ende",
+  tags: ~w(meta ai),
+  description: "Why I built a separate personal blog, and what it has to do with drinking from the firehose.",
+  published: true
+}
+---
+
+I wrote about [publishing short posts](https://www.qwan.eu/2025/05/20/publish-short-posts.html) on the QWAN blog last year. Giving myself license to write shorter, rougher pieces. That worked for a while. But some things don't belong on a consultancy blog.
+
+When I prototype with a coding agent at 11pm, the thing I learn is not a polished QWAN insight. It's a half-formed observation about evals, or a trick for keeping the human in the loop, or just "I built this and here's what surprised me." The QWAN blog has a certain standard. This stuff needs somewhere scruffier to land.
+
+Hence **Firehose** — named after what it feels like to work with AI coding agents. You're drinking from a firehose of generated code, suggestions, and decisions. The interesting question is not "how do I generate more" but "how do I stay in control of what's coming out."
+
+That's also what this site is built with, by the way. The homepage, the blog engine, the layout — all built in conversation with Claude Code. I wanted to experience what our clients experience: shipping something real with an AI agent, and noticing where the friction is.
+
+A few things I noticed, building this:
+
+- **Layout inheritance is a design decision.** The blog engine rendered pages outside Phoenix's layout pipeline. Getting navbar and CSS onto blog pages meant rethinking how the pieces fit together — not just adding a wrapper div.
+- **Warm aesthetics take intention.** The default Phoenix boilerplate is fine, but it says nothing about who you are. Choosing fonts and colours forced me to think about what "personal but professional" looks like.
+- **It's fast when it works, and confusing when it doesn't.** When the agent understands your stack, you move at extraordinary speed. When it doesn't (say, the difference between `@inner_block` and `@inner_content` in Phoenix layouts), you can burn time on a misunderstanding that a human would catch in seconds.
+
+This is the space I want to write in. Shorter than a conference talk, longer than a LinkedIn post. Honest about what works and what doesn't.
+
+If you're a CTO or engineering lead wondering what "AI-native development" actually looks like day to day — not the vendor pitch, the lived experience — that's what I'll be writing about here.