Blog

What these 60 posts are

Nearly every write-up here starts with something on one of my own sites behaving in a way I could not explain, and ends with a number read off a terminal rather than an opinion. 13 of the 60 posts carry a measurement or analytics tag in their own frontmatter, and several exist because the measurement contradicted what I had assumed: the one on content length went back through git to find what five already-ranking pages actually weighed on the day they first ranked, and the order of events came out the reverse of the usual advice. Where something could not be established, the post says so — “What I am not claiming” and “Honest limits” are real section headings in here, not decoration.

The oldest entries, from 2025, are short release notes for things I had just shipped, and they read like it. The long investigations are all from 2026.

The subjects that keep coming back

Counted from the posts' own frontmatter tags. The groups overlap on purpose — a post about whether a sitemap edit moved anything is tagged for both search and measurement and is counted in both — so they add up to more than the collection holds.

• Claude Code, one command per page — 20 posts tagged Claude Code; the ones below drew the most Google Search impressions when this list was picked. How to Stop Claude Code gives the keys that stop a turn, cancel a running tool call and quit, then the parts that surprise people: a Bash command that hits its timeout is moved to the background rather than stopped, and /rewind cannot undo what a shell command did. Make Your Terminal Glow When Claude Code Needs You lights up the right terminal the moment a session is waiting on you, built on hooks, the process tree and a handful of OSC escape codes. Claude Code Notification When Done is a pair of hooks and one short bash script for a desktop notification and a sound whenever Claude finishes or needs your permission, on Linux or macOS, with no jq and no python.
• ffmpeg, one measured command per page — 17 posts tagged ffmpeg or video, led here by the ones that drew the most Google Search impressions when this list was picked. How to Turn a 16:9 Video Into a 9:16 Vertical Short With ffmpeg turns a widescreen clip into a 1080x1920 vertical video by centre crop, blurred fill and letterbox pad, all run on the same off-centre test clip, with a frame from each output and the ffprobe check that proves the file is right. How to Duck Music Under a Voice-Over in ffmpeg lowers background music whenever someone speaks: the defaults only pull it down 5.6 dB under a normal voice, and a measured table of threshold, ratio, attack and release ends on a starting point that ducks by 17.0 dB. We Dropped ffmpeg.wasm for WebCodecs — Here's What That Actually Cost is the one where ffmpeg itself was the problem: the pinned @ffmpeg/core builds turned out to be GPL-configured, including the MIT-labelled one, and swapping in WebCodecs plus a vendored MIT muxer moved jobs FFmpeg had been doing onto our own code.
• Autonomous coding agents, and what they get wrong — 23 posts, none of them prompt tips. What Actually Breaks When a Coding Agent Runs Unattended for Weeks is after-the-fact accounting for twenty-five days of unattended running, in which four of the runs that died mid-task were killed by the wording of the task rather than by the model or the token budget. How to Stop an AI Agent From Working on Itself is the failure mode underneath that: 106 of 187 verified tasks turned out to be the system repairing its own output or its own test machinery. Your Work Queue Can Lie to You takes apart a scheduler that kept reporting an empty backlog while the backlog sat on the board. How to Run a Cheap Model Inside a Coding Agent Loop (and When Not To) is on splitting work between an expensive model and a cheap one held to test vectors it may not edit; the hands-on Claude Code how-tos are in their own item above.
• Search, sitemaps, and whether any of it works — 6 posts. Who Actually Sends Traffic to a New Site? Seven Days of Referrers, Measured is the one that surprised me most: in one week, chatgpt.com sent 111 referrals to Google's 8, and the next-biggest source after ChatGPT was DuckDuckGo, which is mostly Bing wearing a different coat. Does IndexNow Do Anything? I Submitted 2,194 URLs and Measured What I Could kept every HTTP receipt from a bulk URL submission across five domains and then had to publish the awkward part, which is that the one search instrument I own reports on the major engine that does not take part in the protocol. Sitemap lastmod Without Lying and Why Your Sitemap Lists URLs Your Own Pages Disown are the mechanical pair: how to date a page by a signature of its visible text instead of by an mtime a static build resets on every file, and the ways an Astro build ends up listing URLs its own canonical tags disown — trailing slashes, off-origin canonicals, and files that are not routes.
• Publishing code that other people install — 3 posts. npm Resolves Your README's Relative Links Against the repository Field — and a Published Version Can Never Be Repaired is the one I wish I had read first: leave the repository field out of package.json and every relative link in your README resolves against npmjs.com instead of your repo, and because a published npm version is immutable, a listing that shipped wrong can be superseded but never repaired. Cloudflare Email Obfuscation Silently Rewrites @scope/pkg@version in Your Published HTML is a stranger failure — Email Address Obfuscation is on by default and rewrites anything email-shaped at the edge, after your origin has already served correct bytes, and a scoped package identifier with a version pin is email-shaped. 20 Zero-Dependency Libraries, 19 of Them on npm is the catalogue itself: small, tested packages with no dependencies at all — geodesy, solar and lunar maths, climate and weather indices, and everyday CSV, CIDR and SemVer utilities — all but one of them published on a single npm scope.
• Things that run on hardware in the room — 3 posts, every one of them tagged Raspberry Pi. Bluetooth RSSI to Distance gives the formula that turns a Bluetooth RSSI into metres, 126 real readings from a Raspberry Pi 5 scan showing how much the signal jitters, what a 5-reading median does to them, and a standard-library Python script that fits TxPower and the path-loss exponent from your own measurements. YouTube Shorts Automation is a pipeline that generates and uploads a daily short across 5 channels, using AI for script, voice, music, sound effects and video; its write-up opens with the parts list and closes on scheduling and the reliability guards. Phone System is a self-hosted SMS and call gateway with call management and SIM storage.

Where to start

If you arrived from a search, this is the shortest route from the thing you asked to the post that answers it.

• Automating a video channel from a single-board computer — YouTube Shorts Automation. A Raspberry Pi 5, 5 channels, one upload a day, and — unusually for this kind of write-up — the hardware and account prerequisites stated at the top rather than discovered halfway down.
• Finding cut-off text on a page, when the standard advice does not work — scrollWidth Does Not Detect Clipping. The usual probe is scrollWidth > clientWidth. Walked across all 1616 elements of one ordinary page it returned 240 hits and not one of them was cutting anything off. The post explains the ways that reading lies, gives a replacement probe that scored 0 on the same page and still fires on a box that genuinely clips, and then lists what neither probe can see.
• What a month of unattended AI coding actually costs — What Does a Month of Autonomous Coding Agents Cost? My Ledger. Per-run token accounting for all 937 runs: 106.8M tokens, 1,128 shipped tasks, and no revenue. The useful part is not the total but the denominator — and the four consecutive runs that spent 437,315 tokens discovering a typo.
• Checking an invoice against EN 16931 without uploading it anywhere — An E-Invoice Conformance Engine That Reads PDFs. A validator differential-tested against the official Schematron, covering Germany's XRechnung layer, emitting SARIF so a CI job can gate on it, and reading the XML embedded inside Factur-X and ZUGFeRD PDFs. It carries an “Honest Limits” section for what it will not settle.
• Whether writing more actually helps — Does Longer Content Rank Better? I Tripled the Word Count on Five Ranking Pages and Measured. Five pages that already earned Google referrals, tripled in word count, then checked against git for what each weighed on the day it first ranked. The method outlasts the conclusion: it is a few lines of git, and it runs on your pages as well as mine.