Progress notes and build logs from the personal projects I am working on. What I am shipping, what broke, and what I learned along the way.
These are my devlogs: ongoing write-ups about the side projects and experiments I build in my own time. Expect decisions, tradeoffs, dead ends, and the occasional win, documented as the work happens. Use the search to jump to a project or filter by category.
The design for letting Claude search the catalogue, show a movement, write a programme and log a set. Read-only tools first, an identity problem the research did not anticipate, and why the API key trap from the migration applies inside every tool too.
Dropping a Postgres container by moving rows, 2,648 files and every account into Appwrite. Two data shapes that did not survive the move intact, three query shapes only a live instance will tell you about, and a config value that was frozen into the production build at http://build-only.
The board was a catalogue with statistics bolted on top. Rebuilding it as a progression view: a HUD strip, a season grid, milestone tiers, and a day streak that survives a rest day on purpose. Plus the dark mode that existed in the tokens for weeks and was never once reachable.
The board was laid out for a desktop browser that happened to be narrow. Turning it into something usable one-handed in a gym: bottom nav, a manifest, twelve iOS splash sizes, and three bugs a desktop will never show you. One of them was my own service worker stalling every image after the first.
Turning a 1,324-exercise dataset into a gym scoreboard. Why sets are one row each stamped with a local calendar day, why the streak counts weeks rather than days, and the ownership bug that showed the admin 21.9 tonnes of training when only 11.0 of it was theirs.
A process for porting WordPress sites by measuring them, turned into a Claude Code plugin. The one line that tied it to Laravel was not the Blade templates, it was a same-origin iframe. Plus two ways a headless browser hangs forever, a mirror that kept phoning home, and a YAML bracket that silently deleted a command.
Pasting a documentation page into Claude Code means pasting a sidebar, a theme toggle and a footer. The docs site now serves a second channel: an /llms.txt index and a .md twin of every component page, generated from the same registry the HTML reads. Plus the route that quietly ate its own suffix.
The registry goes from 25 components to 35. A toggle that is a checkbox, a slider that is a range input, a hover card that is pure CSS, and a carousel that keeps working with its script deleted. Plus the three bugs that only turned up once I stopped reading the markup and started clicking it.
Every card in the registry drew a black border, and every modal opened with a backdrop that was not there. Both bugs lived in the theme rather than the components, both passed 105 tests, and in the second one the computed style was correct while the screen was blank. What Tailwind v4 changed, and why a colour function can parse and still refuse to paint.
The release was tagged, the tests passed, and the live site was still broken. A gitignore line had frozen the deployed CSS eleven commits earlier, the server was running code from before the merge, and my own syntax highlighting took every documentation page down with a 500. What each failure looked like from the outside, and why the last one was the worst kind of bug.
Every navbar block shipped a full demo page: a min-h-screen wrapper and a hero section that had nothing to do with the navbar. This pass strips them to the header alone, drops Alpine from the docs site itself so it stops proving the opposite of what it documents, and moves syntax highlighting to the server. Including the component attribute that quietly broke the page around it.
Every interactive component in the registry used Alpine.js, which meant copying one Blade file into a fresh Laravel app was never actually one step. 0.3.0 removes Alpine entirely: each component ships a small inline script in the same file, and several of them now need no script at all. Plus the four traps I hit doing it, including a modal that opened invisible.
A 100-page audit wrote a 5.9MB Markdown note, because the whole report was inlined as JSON inside it. Splitting that into three files per site fixed the lag, and rebuilding the dashboard on top of it turned up a CSS specificity bug that had been quietly overriding every button rule I wrote.
The audit engine could tell you a check failed and nothing else, so there was nowhere to put a number. 1.0.2 adds a measurements channel, a streaming API so you can watch a crawl work, and a documentation site. Plus the head-parsing bug that made my social card invisible.
How my Figma to code plugin turns several frames of one page into a single responsive HTML file: name-matched merging, mobile-first breakpoints read from the frame widths, and a mixed-text bug that shipped invisible numbers.
How I added a Tailwind mode to my Figma to code plugin: a v4 browser CDN, faithful arbitrary-value utilities, and a flexbox shrink bug that only Tailwind surfaced.
The second pass on my Figma to code plugin: real HTML tags chosen without AI, mixed text as spans, and sections that reach the screen edges.
How I built Design to HTML, a Figma plugin that turns a selected frame into clean HTML and CSS, and the per-side border bug that taught me to read Figma's stroke model properly.
I ported my ImageMagick VS Code extension to Obsidian. The port removed about 700 lines, and then a rotate slider that would not slide taught me where the real cost was hiding.
A cleanup release for my ImageMagick VS Code extension. No new features, 98 fewer lines, and one real bug that fell out when I deleted the code that was hiding it.
0.4.0 of my ImageMagick VS Code extension adds saved optimization presets and a one-click optimize from the Explorer, encoded in the extension host with no panel and no uploads.
The 0.3.0 release of my ImageMagick VS Code extension fixes drag-and-drop from the Explorer, plus what is coming next.
How I built SafiCSS, a tiny runtime CSS-in-JS library for plain HTML that turns a style object into a class name, with no build step.
Cleaning up vault-crm's Obsidian community review warnings, a dependency override that was quietly broken, and a release that failed on the wrong lockfile.
How I built S-Calc, an Obsidian plugin that evaluates each line of an s-calc code block and prints the answer inline in your theme accent colour. The build story: the engine, live currency, the CodeMirror editor layer, and the 0.6.0 release.
How I built an open take on closed site auditors: 93 checks across 15 categories, a fast crawler, a themed HTML report, and an npm SDK. One runtime dependency at the core.
A minimal Chrome extension that shares the current page to 10 networks or copies it as a blog-ready Markdown snippet, built on just activeTab and scripting so it never touches your browsing history.
Part 2: building the client's lead automation in n8n, wiring automatic email delivery and a Listmonk newsletter sync with double opt-in, then self-hosting it live on their own server.
Part 1 of building a client's LinkedIn lead automation: what the project set out to do, and the information and secure access I gathered before writing any of it.
Build logs and progress notes from my personal projects. Each post tracks what I am working on, the decisions I made, what broke, and what I learned. They are written as the work happens, not polished after the fact.
Whenever a project moves far enough to be worth a write-up. There is no fixed schedule; the cadence follows the work.
Yes. The blog is tutorials and deep-dives meant to teach. Devlogs are running journals of projects in progress, closer to building in public than a finished guide.
Reach out through the contact page. If a project is open to feedback or collaboration, I will say so in the post.