Chameth.com - posts like docker-automatic-nginx-proxy, infinite-avatars, understanding-docker-volume-mounts, why-you-should-be-using-https but not debugging-beyond-the-debugger

Building a family of single-user services

2026-08-27T00:00:00Z

Last month I made and wrote about a personal service to track my workouts and record personal bests. Since then, things have escalated. I now have eight different services running in the same “family”:

Name	Purpose
Collegiate Room	Simple weather overview for my current location
German Mall	LLM broker for use by other services
Gloss White	Weight tracking
Olive Crosses	Central service hub
Pompei Band	Workout and personal best tracking
Speed Carp	Spaced repetition system
Tiger Island	Reading history and library tracking
Typical Horse	Location history and analytics

The names are all anagrams of something related to their function. Finding those anagrams brings me an unreasonable amount of joy. Unscrambling them is an exercise left to the reader. On top of those eight, I currently have four other service ideas sat in my todo list. It’s going to be a big family.

A postmortem on the previous hodge-podge

Most of this functionality isn’t new to me. When I started out playing around with LLMs, I created a “personal data aggregator” that just kept getting things bolted on. I said at the time:

My biggest just-for-me project is a web app that I started to help me aggregate film recommendations. It’s since morphed into a general personal data aggregation service: it deals with data from GitHub, Todoist, Letterboxd, TMDB, Healthkit, and others. It also lets me make re-orderable lists, store recipes, and more. Parts of this could definitely be open sourced, and I might carve them out at some point, but it’s mostly a glorious hodge-podge of things specific to me. Having all these services in one place lets me make quick and dirty automations, for example: when I create a Todoist note on my phone or watch, I often forget to set the due date, so it doesn’t show up in the “Today” view. It was literally a few lines of code to plumb things together so any inbox task without a due date gets set to today automatically.

In a twist that nobody could see coming, it turns out that a “glorious hodge-podge” is actually quite hard to maintain and very annoying to work with. It didn’t help that it had a fairly extensive Svelte frontend and I really am not a fan of JavaScript. I ended up using it less and less just because it wasn’t fun to use or work on, and I didn’t want to waste the limited LLM tokens on doing massive refactors or reworks.

I’ve got almost two decades of professional software development experience under my belt, so you’d think I would have learnt the lessons on how to write maintainable software. Hell, my CV even mentions that as something I’m particularly good at. So what went wrong?

There were a few factors. One of the obvious things to blame is that it was written mostly with LLMs. Most of that blame isn’t very honest, though. In the year or so since, LLMs have definitely got better at writing good code; but I was manually reviewing all the code at the time, and telling it quite carefully what to implement and how to do it. So the buck really stops with me.

I think it’s actually more because it was my first larger project using LLMs. It’s a bit like learning a new programming language. I don’t think the first sizable project I’ve written in any language has ever been any good. You can read and learn as much as you want, but you need to actually run into some problems headfirst to get a grip on how the language works. Or at least I do. LLM coding agents are basically the same deal.

What the LLMs actually did was mask several problems. The architecture wasn’t right for how the service developed. It started out as a single-purpose app and gradually got things bolted on, without any major changes. Everything ended up in horizontal slices: a massive database layer with every query used by every part of the app, a massive HTTP layer, and so on. Working on a single feature became a tap dance all over the source tree. This sort of friction is normally quite apparent when you’re writing code by hand, and at some point it hurts enough that you throw in the towel and do some refactoring. But the LLM will happily plod on without exposing that pain, so the project needed a more active kind of architecture management that it just didn’t get.

The other problem they covered up was how brittle and unmaintainable the JavaScript frontend was. I’m not even sure how I ended up with a Svelte frontend in hindsight. It’s my preferred framework, but I still only reach for it as a last resort when there’s no other way to avoid using one. It seems like there were lots of ways to avoid it here, but I seem to have… well… avoided them. Anyway, each feature got its own slew of Svelte components, without much thought to reusability, shared patterns, or anything sensible. Some of them were just thin wrappers around data supplied by Go, others had vast swathes of logic in them, with no real rhyme or reason as to which way that would fall. It’s the same issue as before: a human would look at it, say “WTF?” and suggest tidying it up; instead, the LLM just gets on with what you tell it to do.

There’s also a big non-LLM factor: one of the primary reasons all of this unrelated functionality got tossed into the same app was that I’d previously worked on a few projects with a microservice architecture and really didn’t like it. My conclusion was they only really worked in large environments where you can dedicate people or teams to working on all the boilerplate that happens from maintaining lots of interconnected projects. Dumping everything and the kitchen sink into one badly architected monolith was, perhaps, not the best rebound from that.

Pompei Band and beyond

When I wanted to implement ~~Strava at home~~ my PB tracking app, it seemed fairly natural to build it out as a new service, rather than bolt it on to the slowly decomposing hodge-podge that was giving off a funny smell. Then when I wanted a weather dashboard, why not do it in the same style as Pompei Band? The same serving architecture, the same page layout… I could even come up with a similar weird two-word name!

Then wouldn’t it be nice if the weather dashboard knew where I was, so it could show the local weather? I could just build that functionality in, but it’d also be a good opportunity to carve out the location history feature from my old service and make a new, shinier one. Then basically every time I even thought about going near the old service, I mentally sketched out a replacement service.

Olive Crosses showing links to the other services

At one point I had the thought that it’d be nice if the services could all link to one another, but I didn’t want to have to touch every service whenever I added a new one. So I made a “hub” service, which maintains the One True Catalogue, and provides a bit of JavaScript the others can all embed to add a navigation sidebar.

So after complaining about microservices and JavaScript, it sounds like I have landed on microservices (or at least miniservices) and JavaScript? Sort of yes, but sort of no. Each service is still its own thing and makes sense independent of the others. There are some cross-service dependencies, but they’re very simple and don’t have the same cognitive or boilerplate overhead as a login service that talks to an authentication service that queries a user manager and so on. And as for the JavaScript, it’s almost exclusively used for progressive enhancement and a few limited client-side visual things like maps and graphs. All of the logic lives in Go, which is a lot easier to maintain and test, and all of the pages are simple Go templates.

Having these services built in a way that I actually like has motivated me to play around a lot more with new features:

My workout tracking app pulls photos from my private Immich server that were taken during workouts, and shows them alongside. Now I can see all my obligatory end-of-run selfies in one place.
The location history app can detect trips away from home, and get an LLM to automatically title them. My most recent, very exotic, trip has the title “A fortnight in Milton Keynes and Coventry”. Each trip gets its own little map, and in the future I want to pull in images like with workouts.
My weight tracking app now exports weigh-ins to Garmin, as it doesn’t fully integrate with Apple Health (it will export its own weights, but not import others, like the ones from my Withings scales). Stupid walled gardens, but thankfully I can fix it myself.
My book tracking app automatically cleans up the titles of books from the nonsense provided by Amazon to something more palatable. “Killing Eve: Long Shot: A twist on obsession, betrayal and revenge from Luke Jennings, author of the TV sensation Killing Eve” becomes “Long Shot (Killing Eve #5)”. It does this by passing the title to an LLM with a bunch of rules.
The hub generates me a snippet of Tailscale configuration that defines the minimal grants needed to allow the intra-service dependencies to work.

The LLM integrations were inspired by my friend Shane who recently demonstrated an inventory tracking app he’d made, which used Haiku to process images of boxes and describe the contents. I’ve dabbled with LLM integration before, and always disliked the process of fiddling with the prompts, and debugging when the LLM gave a nonsensical response. So I built German Mall, which makes it a lot easier. It lets me define a “workflow” consisting of a system prompt, initial message, and expected response format. I can browse past runs, tweak the prompt, rerun them to see the change, all without touching the services using them. Quick iteration loops are the best. German Mall also provides markdown documentation for each workflow, so integrating it downstream is simply a case of pointing a coding agent at the right documentation endpoint.

The “tidy book titles” workflow in German Mall

It’s not all fun and games: having a bunch of services all following the same pattern means you need to do cross-cutting maintenance on them now and then. But that’s surprisingly easy with LLMs. I put all the projects in a common folder, and I launch cross-cutting tasks from that folder. Usually I’ll make the change in one project, and then the cross-cutting task is simply “Look at commit XYZ in ./foo, make the same change to the other projects”.

I’ve also ended up with a shared library used by all the projects to cut down on some boilerplate. That’s normally something I’d resist, especially with personal projects, just because the extra overhead of working in a separate project, managing releases of it, and so on, is a bit annoying. But, again, LLMs can do almost all of that grunt work. They effectively reduce the “cost” of maintaining separate services to a negligible level, leaving me to focus on the more fun aspects like coming up with new names.

The Case of the Unchanging Config

2026-04-09T00:00:00Z

Last week I was attempting to make it so I could share pictures on IRC directly from my client. This sounds simple, but it involves a bouncer¹ that proxies the request to a standalone image hosting service that I had to modify to be compatible. At one point my testing loop was:

Make a change to the hosting service
Commit it
Tag a new release
Wait for it to build
Update the version I’m running on my server
Reconfigure the bouncer to pass the new parameter or change the URL or whatever
Try uploading a photo from my phone
Realise I’ve overlooked something and go back to step one

I could have set up a local copy of everything and tested it sensibly. I should have set up a local copy of everything and tested it sensibly. But it seemed like such a trivial change, and setting up the whole environment seemed like such a pain. After the third or so iteration of failure I was pretty annoyed with myself, computers, and basically everything.

My biggest annoyance was that my bouncer would not pick up the new URL from the config when I changed it. It’s meant to reload the config when it receives a SIGHUP, and it claimed to in the logs, but I could clearly see it was still hitting the old URL. Restarting the bouncer to update the config is a pain, as it disconnects me from all the IRC networks, and has to reconnect to them all, reauthenticate, etc. It also mildly spams everyone who shares a channel with me².

When I finally got everything working I had a look at the bouncer source, and thought I’d spotted the issue. I raised a bug report, ending in this remark:

It looks like the config is reloaded properly, but the handler for uploads is created once at startup and has its own copy of the uploader, so effectively snapshots the config to whatever it is at startup:

fileUploadHandler := http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
    cfg := srv.Config()
    h := fileupload.Handler{
        Uploader:    cfg.FileUploader,
        DB:          db,
        Auth:        cfg.Auth,
        HTTPOrigins: cfg.HTTPOrigins,
    }
    h.ServeHTTP(w, r)
})

I’m not quite sure what I thought I saw there, but that’s not right. I want to call it a hallucination, but that feels like a very overloaded word these days. I guess I was just seeing what I wanted to see, and that was a good excuse to stop investigating. Still, even if I misidentified the cause, the bug was still present, right? … Right?

The Science of Deduction

When the developer got back and said he couldn’t reproduce my issue, I went back to my install and immediately reproduced it. With a calmer head, I figured it was probably something with my particular set up. First thing to check: am I actually running the version I think I am?

Unfortunately the bouncer doesn’t actually seem to expose the version anywhere that I can see. It’s not in the logs, it doesn’t have a -version flag, and none of the IRC-based status commands seem to include it. But I know it’s a Go app, and I know Go embeds the version information. go version -m <binary> will dump it all out, but the binary is inside a Docker image, and the Docker image is one of my nice, minimal, artisanal ones so doesn’t ship a go binary. No problem, docker compose cp bouncer:/bnc ./bnc yoinks the binary out of the container, and then dumping the version shows that, yes, I am running the version I thought I was. Hmm.

The next tool I reached for in my toolbox was strace. Maybe it’s not actually reading the file for some reason? I immediately executed strace -p <pid> -e trace=openat,open,read,pread64 -f without having to look any part of that up. Yep. Definitely. Then I edited the config, HUP’d the bouncer, and saw that it was… reading the config file. As it’s meant to. By default strace truncates strings to 32 bytes, so I couldn’t actually see the line I’d changed. Some more definitely-not-RTFMing later, and rerunning it with an extra -s 65536 let me see the full config. Surprise! The config hadn’t changed!

To confirm my findings, I used docker compose cp again, this time yoinking the config file from inside the container. The inside config file was definitely different to the outside config file. What? My hypothesis at this point was “something something Docker nonsense”. I mount the config as read-only, and was wondering if that meant that Docker was doing something weird instead of just bind mounting it. A quick trip to /proc/<pid>/mounts showed that it was, in fact, not doing anything weird, and was just bind mounting it.

If the file is bind mounted, then surely it’s the same file? I ran stat on the file on the host, noted the inode number, then pondered how to actually run stat on the file inside the container, given the aforementioned awkwardly minimal image. The solution was easy: access it via /proc/<pid>/root/. I could’ve saved myself a bunch of docker compose cp if I’d thought about that earlier. Oh well. The inode of that file was different. What?

The answer was DNS. It’s always DNS. Oh, sorry, force of habit. I meant the answer was vim. I was editing the config in vim, and when it saves files, by default, it writes the new content to a temporary file and does an atomic rename. That’s normally a good thing: it prevents corruption if the write fails midway through³. In this case, though, that means the new file has a new inode. Bind mounting a file binds to the inode, so the container just keeps clutching onto the original config from when it was started, blissfully unaware that the party has relocated down the street.

This problem is likely to happen whenever you bind mount a file into a container. When I mentioned this to a friend, he immediately responded “oh yeah, never do that”, and went on to describe the horrible hacks he’s had to add to Ansible to sidestep the issue. The nicer solution is to just bind mount an entire directory if you can, as then it doesn’t matter what happens to the files within it. I really like having the config files sat alongside the Docker compose files, though; having to create a directory just to work around some bind mount weirdness upsets me.

Now I knew what the problem was, I found there was an issue raised against Docker twelve years ago⁴. The first response was “that’s expected”, and I fully agree with the author: “respectfully, that might be expected by you, but it was not expected by me”!⁵ As a result of the issue they added a nice note to the docs:

Note: Many tools used to edit files including vi and sed --in-place may result in an inode change. Since Docker v1.1.0, this will produce an error such as “sed: cannot rename ./sedKdJ9Dy: Device or resource busy”. In the case where you want to edit the mounted file, it is often easiest to instead mount the parent directory.

But in the intervening twelve years, both the note and the functionality described have gone missing. Ho-hum.

I found it interesting how I’ve only just hit this problem, given how long I’ve used Docker. But I realised that almost everything I run I’m happy to just restart. Cattle, not pets, and so on. My IRC bouncer is one of the few exceptions to that. The only other thing I regularly hot reloaded was Centauri, my reverse proxy, but that had a whole config directory mounted because it was shared between containers, so nicely sidestepped the foot-gun.

So lesson learnt: check for weird bind mount issues before raising issues about config hot reloading. It’ll join the esteemed company of “maybe the drive is full and causing completely unrelated problems?”, “perhaps everything is dog slow because the kernel ran out of entropy?”, and “did systemd sneakily take over that functionality while you weren’t looking?” in the troubleshooting checklist.

An IRC bouncer is basically an always-on proxy. It connects to the IRC networks for you, then your clients connect to your bouncer. The bouncer can then send incoming messages to all your different clients, cache them when you’re offline and replay them later, and lots of other nice things people take for granted in their chat apps these days. ↩︎
Join/part/quit spam is part of IRC, and clients have ways of handling it, but I still feel bad about doing it excessively. ↩︎
You can disable this in vim by setting backupcopy=yes. Good luck redoing all the debugging if you ever accidentally remove that from your vimrc, though! ↩︎
Which is weird, because Docker can’t possibly be that old. That would make me much older than I’m prepared to accept. ↩︎
If you read the thread they didn’t actually mean to sound so dismissive, but it’s still pretty funny. “Yes, it’s expected that the foot-gun causes your foot to hurt. Duh. What did you expect?” ↩︎

An interesting Tailscale + Docker gotcha

2026-01-29T00:00:00Z

As I’ve written about before, I use Tailscale for a lot of things. I thought I had it set up in a reasonably secure manner, but I recently noticed a problem.

I use Tailscale’s ACLs to limit what each node can access, based on the tags I apply to it. So an app node can’t access anything via Tailscale, while an integration or server node can access things tagged with either app or integration. This is expressed pretty simply in the Tailscale ACL JSON:

{
	"tagOwners": {
		// Servers that can be SSH'd into
		"tag:server": [],
		// Applications that are exposed on tailscale but never connect out
		"tag:app": [],
		// Things which talk to other services over tailscale (connecting to apps etc)
		"tag:integration": [],
	},
	"grants": [
		{
			// Users can access everything
			"src": ["autogroup:member"],
			"dst": ["*"],
			"ip":  ["*"],
		},
		{
			// Servers and integrations can access integrations and apps
			"src": ["tag:server", "tag:integration"],
			"dst": ["tag:integration", "tag:app"],
			"ip":  ["*"],
		},
	],
}

Over the past week I’ve been setting up a private Forgejo instance behind Tailscale, complete with an actions runner that runs things using a docker-in-docker container. I didn’t want the runner knowing anything about Tailscale, so I had it configured to speak to Forgejo direct over HTTP (forgejo:3000) instead of using full Tailscale HTTPS URL that I use when accessing it (http://git.example-net.ts.net/)¹.

Everything was going fine, until I forgot to do that translation… and it worked. My Forgejo action runners could access anything on my tailnet. I run several things on Tailscale that just have authentication turned off, on the basis that only authorised devices can access them. Things like the admin interface for this website. I definitely didn’t intend for any workflow I run on my git server to have access to edit my website!

This wasn’t just limited to Forgejo, either. Any docker container I was running could access the tailnet. In hindsight it’s fairly obvious why: the host is running Tailscale, connected as a node tagged with my server tag. That creates a tailscale0 interface, and automagically sets up iptables rules to route Tailscale traffic over the interface. Docker also automagically sets up iptables rules to bridge traffic, and apparently these two sets of rules interact in such a way that traffic from Docker containers is allowed to route via the tailscale0 interface.

I say it’s fairly obvious in hindsight — there’s no reason why Docker would special case any particular host interface after all — but it still feels pretty surprising. Because both bits of software inject their own iptables rules, I never really had a good mental model for how they interact. The host Tailscale node was a completely separate building block to Docker. It would be a pain to use either of them if they didn’t do these rules, but it’s also one of the reasons I don’t really like “magical” things².

There are lots of ways to fix this, but none of them feel particularly great. You can configure both Tailscale and Docker to not automatically fiddle with iptables and handle the rules yourself, but I really hate dealing with iptables³. As a stopgap I did hold my nose and add some iptables rules to drop traffic to the tailscale0 interface if it originated from the IP ranges that Docker was configured to use:

iptables -I DOCKER-USER -s 192.168.0.0/16 -o tailscale0 -j DROP
iptables -I DOCKER-USER -s 172.17.0.0/12 -o tailscale0 -j DROP

The DOCKER-USER chain is a nice little escape hatch; it comes before the main auto-generated DOCKER chain, and Docker leaves the rules in it alone.

So I added these rules, and felt pretty good about myself, and then… everything started breaking in weird ways. After some debugging I realised the problem was DNS⁴. The server’s DNS resolver is 100.100.100.100, a special Tailscale address. This is so that it can resolve tailnet hostnames via MagicDNS (and so I don’t have to configure my custom DNS servers manually on each device; Tailscale does it for me). My new iptables rules inadvertently dropped all the DNS packets coming from docker containers. D’oh.

Obviously the solution here is to double down and add MORE iptables rules:

iptables -I DOCKER-USER -s 192.168.0.0/16 -d 100.100.100.100 -o tailscale0 -j ACCEPT
iptables -I DOCKER-USER -s 172.17.0.0/12 -d 100.100.100.100 -o tailscale0 -j ACCEPT

This got DNS working again, but it all feels a bit gross. Not least because to actually make these persist you need to use iptables-save, and then you also get all the rules that both Docker and Tailscale have inserted. I came across a script to filter out the Docker ones, but… yuck⁵.

I realised a cleaner idea might just be to lock down what server tagged nodes could access within Tailscale itself. That way I can avoid touching iptables at all. It’s handy to be able to curl services from the host when debugging, but it’s not really necessary. So I removed that access, and… stuff broke again. I’m using my Forgejo instance as a registry for some of the docker images I run, so the docker daemon needs to be able to reach it. I ended up making a new tag for infrastructure, which can be accessed from server devices. This does still allow all the docker containers to reach Forgejo, but I already have it set up with appropriate access controls and public/private repository splits. Forgejo is a service designed to run publicly, so this seems a reasonable trade-off for convenience. I used the tests feature of Tailscale’s ACL config to make sure I’d got the rules right:

{
	"tests": [
		{
			// Servers can only access infrastructure
			"src":   "tag:server",
			"proto": "tcp",
			"allow": [
				"tag:infrastructure:8080",
			],
			"deny": [
				"100.84.16.43:8080",
				"me@example.com:8080",
				"tag:server:8080",
				"tag:app:8080",
				"tag:integration:8080",
			],
		},
	],
}

The whole situation still feels a bit messy. If I ever get around to switching to nftables I might loop back and manually craft some rules for routing traffic, instead of leaving Tailscale and Docker to do their own thing.

I discovered later on that logging into the container registry at forgejo:3000 actually issued a redirect to https://git.example-net.ts.net/ so this was all basically for naught… ↩︎
I didn’t use Tailscale’s MagicDNS for a long time just because the word “magic” put me off. Only when I eventually got around to learning how it worked, and seeing that it wasn’t really that magical under the hood, did I change my mind. ↩︎
nftables seems far better in lots of ways, but I can’t really be bothered migrating. Maybe next time I reimage the server for whatever reason… ↩︎
It’s always DNS… ↩︎
Again, nftables would almost certainly help here. It actually has (gasp) configuration files. But again, I really didn’t want to spend the time migrating. ↩︎

Making a font of my handwriting

2025-08-08T00:00:00Z

Recently I’ve been on a small campaign to try to make my personal website more… personal. Little ways to make it obvious it’s mine and personal, not just another piece of the boring corporate dystopia that is most of the web these days. I don’t quite want to fully regress to the Geocities era and fill the screen with animated under construction GIFs, but I do want to capture some of that vibe.

I’d added some bits and pieces along those lines: floating images in articles now look like they’re stuck to the page with sellotape, related post links have a wavy border that animates when you hover over them, and so on. Next, I wanted to change the heading fonts from a monospace font to something cursive, to resemble handwriting. Less terminal output, more handwritten letter. I couldn’t find one I liked, though. So why not make my own? It can’t be that hard, right?

Failing to do it myself

I set out to try to make the font myself using open source tools. After doing a bit of research, it seemed like the general approach was to create vectors of each character and then import them into a font editor. That seems to mean either Adobe Illustrator and FontLab (if you have too much money) or Inkscape and FontForge (if you like open source). I fall firmly into the latter category, so I grabbed my graphics tablet and opened Inkscape.

I wrote out my first three letters: capital A, B and C. Saved them in Inkscape, and attempted to import them into FontForge. Then I remembered one crucial thing that had slipped my mind: I absolutely loathe using FontForge. It’s a bit like when you open an old version of GIMP and get a bunch of weird looking windows floating all over the place; it feels like you’re fighting against the tool to do even the most basic operations. The difference is I have cause to edit images a lot more than I edit fonts, and GIMP has actually significantly improved their UI over the years.

Here are the rough steps I went through with FontForge:

Launch Font Forge. It shows a weird bit of art in one window, and an open file dialog in another.
I don’t want to open a file, so I close that dialog. The program exits.
Relaunch Font Forge, and realise that within the “Open Font” dialog is a “New” button. Click it.
Get to the standard font-editing UI. Right-click on the “A” looking for a way to import an SVG. Don’t see one.
Click around a bit, exploring the menus. Everything feels a bit off. You can’t open one menu then hover over the next to see its content, like basically every UI toolkit in existence. I think FontForge has eschewed QT and GTK in favour of doing things itself.
Find the “Import” option in the File menu. Hope it’s for a single glyph not the whole font.
A file picker opens. Again it’s all a bit off from normal desktop conventions. Try to resize it, and just get blank grey space at the bottom.
Type the absolute path I want to go to in the text field.
Get a dialog saying “Not a bdf file /home/chris/etc”. Press OK.
Get a dialog saying “Could not find a bitmap font in”. Press OK.
Press Ctrl+L to see if that lets me enter a path. Click everything in the dialog to try to find a way to enter a path. Get annoyed. Give up. Click through folder-by-folder to get to where I want to be.
Get to the folder and don’t see any files. Change the format to “SVG”. Double-click the newly-visible SVG file.
Get a dialog saying “You must select a glyph before you can import an image into it”. Press OK.
The import dialog goes away, having not imported.
Select the glyph in the main tool area, then repeat the File→Import dance.
It’s actually there now! Open the glyph in the editor and see it’s a complete mess of Bézier curves. I can’t click what I want without accidentally moving a handle for an adjacent curve.
Rage-quit.

I’m sure FontForge is less anger inducing once you’re used to it. And you definitely could use it to build a font like this if you had much more patience than me. I’d had enough of death-by-a-thousand-paper-cuts though.

I briefly tried Inkscape’s built-in support for making an SVG font. It annoyed me a lot less, but it’s fiddly: it seemed like each glyph had to be a single path, so you had to convert the glyphs to paths, then merge them correctly. If you merge them incorrectly then the wrong bits of your letters end up filled (like the inside of the ‘B’). Path manipulation is getting towards the limit of my knowledge of vector editing, and it took a bit of trial and error for each letter that had more than a single stroke. I didn’t fancy doing that for every letter.

I’m usually a big advocate of open source, but this was one of those painful times when it feels like it just falls short. Clunky, painful UI and processes where commercial tools just let you get on with your work.

You can exchange money for goods and services

When I’d been looking for open source tutorials, I found many mentions of a closed source, hosted tool: Calligraphr. It has a free version with limitations (no ligatures, no variations, 75 glyphs per font), and a pro version for £8/month. I’d normally balk at the idea of a subscription for this, but they have the perfect answer: you can make a one-time payment, and your account automatically downgrades back to free after a month. It’s not a hidden option, either, it’s the most prominent button on the upgrade page. That made me happy to give them £8 to play around with the service for a month.

Calligraphr works by having you print templates, write out the letters, then scan them in. It does some magical processing to extract the glyphs, provides tools to tidy them up, align them, etc, and then produces a TTF file for you. You can see some of my completed templates here:

Most of the templates I used for the font

Calligraphr has a nice UI to generate the templates, allowing you to select which glyphs to include. I added the “minimal English”, “basic punctuation” and “Ligatures” sets. That gave me four pages to fill out, and I did them all twice. That let me filter out versions that didn’t work well, and have variants for some letters so the font wasn’t too repetitive. Later on, I went back and added some custom ligatures based on blog post titles that didn’t look quite right: “Re”, “To”, “ers”, “ey”, “hy”, “ra”, “re” and “ty”. Ligatures like this help it look more natural: when we write we don’t just stamp out identical letters regardless of their surroundings, instead they will connect to their neighbours, or overlap slightly, or even share a stroke.

I filled these templates in with a Sharpie, as I wanted a fairly informal, scrap-booky look, and it would also give good solid shapes that should be easy to pick out of the template. I scanned them with the “Scan Document” function on my iPhone, and uploaded the PDFs to Calligraphr.

Iterating and tweaking

The Calligraphr UI allows you to preview the font, but I found it a lot more useful to just download a copy and use it on a local copy of my website. That let me test it with real text, and see how it’d look at the different font sizes I use on the site.

The first version was not great. Despite the guidelines on the template, I apparently wasn’t good at sticking to them. Some letters were floating way off the baseline, and some were sunken below. When those opposites met it looked terrible. Fortunately Calligraphr has a pretty easy tool to slide each letter up and down, and scale it up or down if needed, and you can see it next to other letters as you do it. It took a little bit of time to go through all the variants of all the letters, but the next version looked a lot better.

Another tweak I ended up doing was reducing the spacing between letters. The defaults Calligraphr uses are probably good for a blocky font, but I wanted to put the letters close together to give it more of a joined-up look. Again, this is an easy tool to use, you just drag the sides in or out as desired. While these tweaking steps were probably as fiddly as some of the Inkscape steps I refused to do earlier, they’re a lot more rewarding as you see things improving with each one. It’s a lot easier for me to commit time and effort to improving something that’s already working reasonably, than put that time and energy into an unknown.

Later, I noticed that occasionally there would be a huge gap in a title. Not “the kerning is slightly off” but “there’s enough room to park a bus”. It took me a while to figure out what was happening: a couple of glyphs hadn’t been isolated perfectly and had picked up a few pixels from the template lines at the edge of their boxes. That meant the glyph had a width that covered the actual written glyph, a big gap, and then the rogue marks. At first, I fixed this by just adjusting the width, but that left the little pixels floating awkwardly down-sentence. The proper fix was to use the editing tool and simply delete them, and then Calligraphr snapped the width back to what it should be.

These iterations took a while to do, but I just dipped in and out occasionally over the course of a week, so it didn’t actually feel like too much work. I quite enjoy the process of refining things, too.

Result and a surprise

If you’re viewing this post on my website¹, you can see the font in the headers, captions, and a few other places. Here’s how it compares to my actual handwriting:

My handwriting vs my handwriting font

It’s not close enough to forge documents, but I think it definitely gets across my style, and that’s exactly what I wanted. It’s surprisingly legible even at smaller font sizes — I think the weight of the Sharpie helps here — and at £8 and a bit of manual work was a lot more economical than spending days wresting with open source tools.

A few weeks after I put the finishing touches on the font, I got an e-mail from Calligraphr. As my account had lapsed back to the free version, I was no longer eligible for the “server-side backup” feature. So what did they do? They e-mailed me an exported copy! It’s a JSON file with the properties of each glyph and a base64 encoded image. Not only can I re-upload this to Calligraphr if I resubscribe, I can probably hook something up to edit it should I ever need to. I’m blown away by how pro-user Calligraphr’s business practices are. They’re up-front about pricing, don’t try to get you stuck on an auto-renewing subscription, and automatically export your data. It’s like a breath of fresh air compared to the barrage of dark patterns that other websites foist on us. If you want to make this kind of font, I’d definitely recommend them just because of how nice they are.

And I haven’t changed everything since writing this post… ↩︎

Coming around on LLMs

2025-05-28T00:00:00Z

Claude says hi.

For a long time I’ve been a sceptic of LLMs and how they’re being used and marketed. I tried ChatGPT when it first launched, and was totally underwhelmed. Don’t get me wrong: I find the technology damn impressive, but I just couldn’t see any use for it.

Recently I’ve seen more and more comments along the lines of “people who criticise LLMs haven’t used the latest models”, and a good number of developers that I respect have said they use coding models in some capacity. So it seemed like it was time to give them another shake.

The first decision to make was which model to try. OpenAI are no longer the only player in the game, every tech company of a certain size is now also somehow an AI company¹. I looked a bit at some benchmarks, and then mostly ignored them and went with the only company that I didn’t outright hate: Anthropic, and their model Claude.

Initial impressions

The latest Claude models do feel a lot more “capable” than the earlier ChatGPT versions I remember, but they also still have a lot of the same problems. At their heart, they’re still text-prediction models, and still seem to be trained to predict text that will please the user rather than be factually accurate or useful.

One of the improvements that people kept mentioning was the ability for models to access the web directly. I’ve seen it kick in a bit naturally, and asked for it explicitly sometimes, and it’s… nothing special? Are people just really bad at searching the web? Maybe they should try Kagi? I feel like about 80% of the time I could have found the information just as quickly myself, about 10% of the time it searched and then hallucinated an answer, and the remaining 10% it found something quicker than I otherwise would have. Those aren’t great results, especially when you consider how much cheaper and simpler just searching the web is.

There is another way to trigger web searches in Claude: using “research mode”. When you enable it, it splits off into different models: a “lead researcher” to come up with a plan, and then some minions that execute it. It ends up doing hundreds of web queries in the span of a few seconds. Suddenly I understand why things like Anubis need to exist. I’ve not used the “please launch a DoS attack” button since.

What did impress me, though, was its ability to churn out reasonable-ish code. It can hack together a bash script as well as I can, and do it far faster than I’d be able to. Sometimes they even work. That made me wonder what it would be like doing actual coding with it. Anthropic have a CLI tool called claude-code, so I paid them lots of money and gave it a spin.

Coding with Claude

The first thing I notice about claude-code is that I really like the interface. It’s basically an input box in a terminal. It doesn’t force me to use a certain IDE or do things in a certain way. By default it asks before making any changes, showing you a side-by-side diff of what it’s doing and allowing you to provide feedback. If I had to design a way to interact with a coding agent from scratch, I can’t think of many things I’d improve.

A simple example of a claude-code session

The power in claude-code versus just using the web UI is that it can use tools. It can query git, run grep commands, even use sed if it wants to change something in lots of files at once. It’s very good at figuring out its way around a codebase, even without any explicit instructions. You can see in the screenshot that with a little prompting it managed to get the permalink to this post; I didn’t tell it where the posts were stored, or how to work out the latest, I just told it when it was wrong. If I’d run /init before it would have probably picked up on the fact that my posts have custom permalinks, and noted it in CLAUDE.md.

But how good is it at actually writing code? It’s like having a keen but not particularly thorough Junior Engineer at your beck and call. If you give it a clearly defined task and guidance on how to implement it (and maybe some feedback as it suggests changes), it’s more than capable of doing it. If you don’t give it enough guidance it tends to go more off the rails. I tried having it generate a simple application from scratch with minimal technical guidance and no review of what it was doing, and it made such a mess of it I decided it was quicker to throw it away and start again by hand.

The man behind the curtain

Even with sufficient guidance, at times it’s really obvious that it’s an LLM generating pleasing-token-strings and not something that genuinely understands what it’s doing. It will spit out code like this:

if err != nil {
    if err == sql.ErrNoRows {
       return nil, err
    }
    return nil, err
}

Why’s that check for sql.ErrNoRows there? It’s entirely pointless. There was no instruction to check for it, none of the existing code checked for it, but I assume it comes up quite a bit in the training data. But it didn’t understand why, so it put the check in, and returned the exact same thing as if it hadn’t.

It also occasionally tries to “cheat” or solve problems the wrong way. It sometimes feels a bit like you’re asking for wishes from a Monkey Paw. “Stop the unit tests failing”, you’ll say; Claude will respond with a request to delete the failing test. The man behind the curtain isn’t particularly well hidden, and the training data and pattern matching often shows through.

This lack of understanding also makes it very hard to get Claude to use comments in a sensible way. It absolutely loves doing nonsense like this:

// render form
function renderForm() {
    // ...
}

Even with explicit instructions to avoid useless comments, or comments that only explain what the code does (not why it does it), or other prompts. Again, I’m chalking it up to a mixture of training data that does that, and lack of any actual understanding about why a human developer may want a comment to exist. Of course, there are plenty of flesh-and-blood devs out there that also don’t comment effectively, so maybe I shouldn’t give Claude too much of a hard time on this.

… and yet …

So with all those problems, it sounds like it’s just not worth it, right? Well, not quite. I feel like programming mostly consists of two distinct tasks: thinking how to implement things; actually implementing them; and then debugging why you’re off-by-one somewhere. Letting an LLM do the thinking is a big no-go for me: it’s evidently not good at it, and frankly that’s one of the things I most like about programming. Having let it loose on small projects, I shudder to think what the codebases of all the “vibe coded” projects that are popping up are like.

For the second stage, though, it’s actually quite nice. If I’ve thought through how I want something to be implemented, I can feed those steps to Claude and have it churn out the otherwise not-too-interesting code. This requires far less time and concentration on my part than writing code, to the extent that I can be thinking about the next feature, or doing something else at the same time. I don’t feel like I’m going to end up deskilling myself this way, as I’ve already formed the idea of what I want to code, I’m just using the LLM to spit it out faster than I can type it.

A lot of the most boring bits of coding like implementing CRUD-y operations can be summarised as “Look at this file/function. Do the same thing but slightly differently elsewhere”. Claude is great at this, especially when explicitly prompted like that. You’re basically playing to its pattern-matching strengths, rather than asking it to come up with anything novel. Most of my prompts tend to be prefixed with “Look at @some_file.go and @other_file.go.” to cue up the patterns I want it to use, and I find this works well.

As for debugging, it’s a mixed bag. It’s sometimes amazingly insightful, and sometimes just runs around in circles trying the wrong things over and over. It’s worth asking the question, but I definitely wouldn’t rely on it over my own abilities.

The future

I’m probably going to carry on using claude-code, at least for personal projects. I’ve got so much done that I just wouldn’t have been bothered to do if I was doing it all by hand. I very much enjoy being in the more “diffuse thinking” mindset, planning how things are going to work, rather than being stuck in the mines digging out SQL queries. After all, who wouldn’t want an over-eager assistant to work on all their hobby projects?

Work is a slightly different matter: a private project or even an open source project that disclaims any liability is different to something I’m being paid to deliver, and bear responsibility for fixing if it’s not done correctly. I’m not saying I won’t use it at all, but if I do it’ll be much more constrained than I would in personal projects.

As for non-code usages: I’m not sold. My sceptic hat is still firmly in place. I don’t think chat is a particularly good interface for many things, and hallucinations are still a big problem despite what people say. I hate the tide of AI slop that’s taking over the Internet, and how LLM-powered chat agents are being forced into every random product. You’re definitely not going to be seeing any AI-generated blog posts from me!

One topic I’ve not gone into here is the ethical concerns about using LLMs. They obviously exist, and I do have thoughts, but that’s a topic for another day.

I’m surprised it’s not gone more mainstream: why’s there no Tesco Value LLM model, yet? ↩︎