experiments with claude, part ⅳ: dzilification of MIME-Lite

Is all my 2026 blogging going to be about doing more random stuff with Claude? No, I promise. But I’m still working through my backlog of “and the next thing I tried to put it through its paces”, so here it comes. This time, more drudgery.

Project Four: MIME-Lite to Dist::Zilla

Here’s a quick one before I post beefy Project Five tomorrow.

A couple weeks ago, I got a pull request against MIME-Lite. That’s a pretty popular Perl email library that I maintain, and which Fastmail uses at least a little. The contributor had supplied a patch to use GitHub Actions to run the tests on every PR. Great!

Thing was, I already have a standard way of doing that, and I didn’t really want to pick up another way. The other thing was, my standard way relies on the distribution being built with Dist::Zilla. It’s not hard to convert from not using Dist::Zilla to using it, but it’s not trivial. And I don’t love MIME::Lite, so I wasn’t keen to spend a bunch of time on it.

I fired up Claude, pretty certain that I was going to get a crappy result.

rjbsConvert this distribution to building with Dist::Zilla, in the fashion of most RJBS-authored distributions, like Email::MIME.

claudeI’ll explore both the current MIME-Lite distribution and Email::MIME as a reference for the Dist::Zilla conversion pattern.

In about 20 minutes, it was done. Now, when I say “it”, I mean Claude. Claude wrote a big commit that wasn’t good enough. MIME-Lite had a bunch of little quirks, like optional dependencies and a slightly not-rjbs-standard documentation structure. I spent just a few minutes polishing things up and it was good to go.

This was a good example of needing some expert involvement. Understanding exactly what to do was going to require way too much context that couldn’t be found in the repository or in “understanding” of published copies of things. But some exploration of MIME-Lite and Email-MIME let Claude do a bunch of the work so I didn’t have to grumble my way through it. As I’ve seen a few times, it also did work I would not have bothered with, like converting the changelog to a more consistent format.

One funny nit… the original changelog had:

=item Version 1.101   (1997/03/01)

Baseline code.
Originally created: 11 December 1996.  Ho ho ho.

The new one:

1.101     1997-03-01
        - Baseline code. Originally created: 11 December 1996.

I guess Claude is part of the war on Christmas.

Claude also hallucinated author names weirdly and screwed up the Copyright. You can’t just let it do whatever it wants with no oversight! It was pretty bad, and I can definitely see somebody getting into hot water this way. This was probably the worst Claude experience I had that didn’t end in scrapping things. I think I probably could’ve done this one myself faster and with fewer errors. As the world’s leading Dist::Zilla expert… probably shouldn’t be a surprise.

With the conversion done, I dropped the standard CI workflow in place and… found problems in my conversion. Just what CI is for. Now it’s all sorted and happy and any new PRs will get test runs.

You can read my transcript of this conversation, but it’s a bit odd. I stopped and resumed the session at one point, and Claude waited at a permission prompt for 12 hours while I slept or something.