Collection / Log

Small signals.

Chronological micro-posts for observations, decisions, and lessons while they are still fresh.

November 13, 2026

One colour per app

I wanted to tell a dozen windows apart at a glance, so I ended up forking a small menu-bar app and giving every app its own outline colour.

Half the windows on my screen are now the same window. Same sidebar, same composer, same empty space where a document would be. Three of them are different tools pointed at different folders, and at a glance they are interchangeable.

The cheap fix is a border. Not the kind a window manager draws, where the outline marks which window has focus. I wanted the other thing. Every window wearing a colour that belongs to its app, so position and hue together tell me what I am looking at before I read a single word.

I found a small app that already worked that way and then spent an evening making it work better, which is a fair description of most of my setup notes.

What the app gets right

window-border is a macOS menu-bar app, MIT licensed, 298 lines of Objective-C in one file. It is a fork of a fork’s worth of small decisions that all line up.

The first is what it refuses to ask for. It reads window bounds and the owning app from CGWindowList, which needs no screen recording permission. Anything that draws over other windows usually pays that toll, and I would rather have a coloured rectangle than a permanent screen capture entitlement.

The second is how it draws. Each target window gets a transparent, click-through overlay of its own, ordered just above that window. The borders stack with the windows instead of floating in a layer above everything, so a full-screen app still covers its neighbours properly.

The third is the boring one that made me trust the code. The overlay window class overrides constrainFrameRect:toScreen: and returns the frame untouched. AppKit’s default behaviour is to nudge a window with no title bar back onto the screen, which would put every outline a few points off the thing it is supposed to outline. Somebody hit that bug, understood it, and wrote one method to stop it.

One point of outline is not enough

The stock border is one point wide, which is enough to know a window exists and not enough to find the active one. So the fork has a settings.json with a width and an activeWidth, the second one applying only to the frontmost window.

Two details in that are worth copying. Both are capped, at 20 points, because a config file is not a promise that a value makes sense. And extra thickness grows outward from the window edge rather than centring on it, so a thick border never creeps over the content. The inner point stays exactly where it was, and the rest of the width pays for itself in desk space.

The same reasoning gave the other two branches. A per-app glow radius for terminal panes that sit side by side and need a softer cue than a hard line, and a "*" key in the colour map so unlisted apps get a colour I chose instead of the next one in an arbitrary palette.

None of this is glamorous. All of it came from using the thing for a week and noticing what was still hard.

The detail that proves someone read the source

The README says the geometry timer runs at about 20 Hz. The source says kTick = 1.0 / 60.0, with a comment reading 60 Hz. Nobody is lying. Someone changed the constant and did not update the sentence, which is how nearly every stale document happens.

I only found it because I wanted to know what the polling cost was before I installed something with a timer in it. Reading the 298 lines took less time than reading a vendor’s feature page.

Where a border stops helping

A colour tells you which window, not what is in it. If two sessions of the same tool are open on two different projects, they get the same colour, and the border is useless. That is a folder and naming problem, and the fix is a window title I set myself or one window per project.

The border solves recognition. Comprehension still comes from how the work is organised, and no amount of outline solves that.

The app itself is in the Toolkit, with the fork and the upstream both linked, so the credit lands where it belongs.

Read log →

October 30, 2026

Repainting the site in a brand palette

The blog now wears a cool graphite brand palette, with a dark theme built from the same ramps and a rollback that takes one command.

I moved the blog onto a brand guide today, cool graphite with a graphite ground. The guide already existed for client work, so the job was making a personal site wear it without losing what a blog needs.

The light theme is the guide applied directly. A graphite ground at #F5F6F7, white surfaces, #C6CBD1 hairlines, petrol #093558 headings, Indigo Action #3548A8 for links, and red held in reserve. Nothing glows and nothing gradients.

Where the decisions live

Two files carry the theme. CustomStyles.astro holds the full token set, copied from the brand repo so the blog and the guide move together, plus the handful of aliases the AstroWind template reads.

The second file does the heavier work. The template ships hundreds of text-gray-900 and dark:bg-slate-900 classes across about forty components, and editing each one would have been a week of churn with no design benefit. Instead tailwind.css re-points the stock Tailwind palettes at the brand ramps. Lightest stays lightest, darkest stays darkest, so every existing utility lands on a brand colour and the dark variants keep their meaning.

Dark mode is an extension, not a copy

The guide defines a light mode only. It says a dark surface is a petrol navy panel inside a light page, which is right for documents and wrong for a site that people read at night.

The dark theme here is graphite night, an inversion built from the same ramps. #0B0C0F ground, #14161A panels, #2C3036 hairlines, Pale Mist headings, Sky links. Same palette, opposite ends. That keeps one brand with two lighting conditions instead of two brands.

Type

Headings use M PLUS Rounded 1c at weight 800, body uses Noto Sans, and Noto Sans JP sits in the stack behind it. Astro’s Fonts API self-hosts and subsets all three, which comes to five woff2 files, and the Japanese ranges are never fetched on an English page.

The rounded terminals change the tone more than I expected. The old Inter made everything look like software documentation. This reads warmer without going soft.

Red stays rare

The guide caps red at five percent of any surface. On the blog that means one element, the filled button, and nothing else. The header has no call to action right now, so the red is wired up and currently invisible. Adding a single entry to the navigation actions would paint it. That is a content decision waiting, not a missing style.

A build is not a screenshot

A passing build says nothing about what a page looks like. Two commands confirmed the code compiled. Neither confirmed the colour.

So I drove the built site in a real browser and read computed styles in both modes, then swept every text element, walked up to the nearest opaque background, and computed the contrast ratio for each pair. Two real problems surfaced.

Article headings were inheriting the body face, so the display font only reached the homepage. A base rule fixed it for every page at once.

Dark code comments sat at three to one against their panel, which is thin for monospace text. Switching the dark syntax theme to the high contrast palette fixed it without touching a single token.

Everything else passed, in both modes, across six page types.

Rolling it back

The theme landed as one commit with a tag on the commit before it. Ten files changed and none of them are content. If the direction turns out wrong, reset to the tag, rebuild, deploy, and the blog is exactly as it was an hour earlier.

That is the part I care about most. A redesign you cannot undo is a redesign you hesitate to start, and hesitating is how a site ends up three years out of date.

Read log →

September 22, 2026

Adding tools and people to the garden

The garden now has a Toolkit for the software and services I use, plus a People map for the thinkers and creators whose work I return to.

I added two new ways to connect the garden.

The Toolkit is a collection of the software and services I use. Each entry links to the tool’s website and can include learning resources and related writing on this site. The index makes it easier to link to a consistent reference whenever a tool comes up in a note or article.

The People map gathers the writers, researchers, and builders whose work shapes my thinking. A profile can include a homepage, useful links, and explicit connections to other people. Notes, articles, and log entries can identify people in their frontmatter, so the map can show where their ideas appear without trying to infer a mention from prose.

Both structures are intentionally simple. They make the garden more useful now while leaving room for richer connections as the writing grows.

Read log →

September 22, 2026

Adopting markdown-style links over wikilinks

Foam and Obsidian are set to write standard relative markdown links instead of [[wikilinks]], so a remark plugin now rewrites them into site routes at build time.

Clicking the log link in an earlier note failed in the editor: it used a root-relative site route like /log/starting-an-astro-site, which Astro’s router resolves but VS Code and Foam try to resolve as a file path.

I write this garden in Foam and Obsidian as well as in the site itself, and both are configured to use plain markdown links ([text](../log/entry.md)) rather than [[wikilinks]], so the same files stay readable on GitHub. That meant the site needed to meet the editors halfway instead of the other way around.

The fix is a remark plugin, gardenLinksRemarkPlugin in frontmatter.ts, wired into the markdown processor in astro.config.ts. It walks each file’s links during the build and rewrites any relative .md/.mdx link that points inside src/data/<collection> into that collection’s route, e.g. ../log/starting-an-astro-site.md becomes /log/starting-an-astro-site. Links that are already absolute, external, or root-relative are left alone.

Now the convention is simple: write relative file links everywhere. They work as file links in Foam, Obsidian, and GitHub, and the build turns them into working site routes.

Read log →

September 22, 2026

Configuring workspace Prettier

VS Code now uses the AstroWind project's local Prettier and its Astro plugin instead of the extension's bundled formatter.

Today I configured VS Code to use the Prettier installation that belongs to this AstroWind project. The workspace setting points the Prettier extension at ./node_modules/prettier, and Astro and Markdown files now use that extension as their default formatter.

The project already had the right Astro formatter setup: prettier and prettier-plugin-astro are development dependencies, and .prettierrc.mjs explicitly registers the Astro plugin and parser. The change was needed because the editor was loading its bundled Prettier, which could not resolve the project plugin from /noop.js.

The site can now format from VS Code and from the command line with the same project-local formatter. Markdown uses Prettier’s built-in Markdown parser; .astro files use the Astro plugin.

Read log →

September 21, 2026

Starting the Astro site

The first setup decision was to define what the Astro site is for before choosing its visual details.

Today I started turning the AstroWind foundation into something that belongs to me. The first useful decision was not a color or a component. It was deciding what the site is for: a place to develop ideas about leadership, technology, operations, and change.

Astro is a good fit because the site can stay close to the content. I can begin with a Markdown file, connect it to a note about the site structure, and later develop the pattern into an article. The site-building map collects the path.

The work feels more manageable when each decision leaves a small, inspectable trace.

Read log →

September 21, 2026

Why I am making this garden

A short introduction to why I am building a digital garden for unfinished and developing ideas.

I am making mikes.blog as a place to put ideas before they are polished. Some thoughts will be useful only for a day. Others will grow into notes, articles, or maps that help me see a subject more clearly.

This is also a record of how I work: curious, practical, systems-minded, and interested in the distance between a good strategy and a useful result. I want this garden to hold that movement in public, one small piece at a time.

Read log →