This blog started writing itself from each day's changesThis blog started writing itself from each day's changes

July 31, 2026 · A post-commit hook now hands each day's git history to a session that writes this. Text-message dispatch was built, then removed the same evening.The site now writes its own daily post from what changed. Sending jobs by text message was built and then taken back out hours later.
last updated

The blog writes itself

This post is a little recursive, because the system it describes is what wrote it. Committing anything now schedules a job for that day; the job reads the day's git history and writes the post you are reading, then publishes it.

The interesting part was not the writing but the triggering. A commit hook that fires per commit would spawn a session per commit, so it debounces on the date rather than the change. A day of forty commits produces one post, not forty. Later the same day the prompt learned to split a multi-initiative day into sections, which is why this one has headings.

Who do I need to text

The day's biggest piece was a reminder feature built on top of the existing CRM rather than beside it. It reads the local message history, works out which conversations end on someone else's message, and can be asked three ways: out loud, on a schedule, or on a private web page where a thread can be dismissed.

The scheduled half runs every twelve hours as a per-user job rather than a system one, specifically so it inherits the logged-in user's permission to read the message store. A system job can never hold that grant, so the choice is forced rather than stylistic.

One assumption from the write-up did not survive: registering the new voice intent had been marked as a step needing a human in a web console. It turned out the command-line tool on this machine was already authenticated and could push the model directly, so the whole thing shipped without leaving the terminal.

Testing it for real then caught a genuine bug: the CRM was deciding group-versus-one-to-one from a field that is unreliable, and had been silently excluding every real conversation on this machine. It now keys off the number of handles in the thread.

SMS, built and unbuilt

The other new way in was meant to be text message: send a job by SMS, no Echo required. It was built, wired up, and then removed the same evening once it turned out the provider has no personal single-sender tier. The cheapest real path was a business account and a compliance review for a service with one user.

Walking a feature back out the day it lands is not a wasted day. The privacy and terms pages it required stay up, because they were worth having anyway, and now describe what the site actually does.

The blog writes itself

This post was written by the thing it is about. Whenever I change anything, a job is scheduled for that day; it reads everything that changed and writes the post, then publishes it. It waits until the end of the day rather than reacting to every single change, so a busy day produces one post instead of forty.

Who do I need to text

The biggest piece was a reminder that tells me who is still waiting on a reply. It reads my messages, finds the conversations I have left hanging, and can be asked out loud, on a schedule, or on a private web page where I can tick things off.

The scheduled version runs as my own user rather than as a background system service, because only my own account is allowed to read my messages. Testing it properly also caught a real bug: it had been quietly ignoring every genuine conversation on this machine.

Texting, built and unbuilt

The other idea was sending jobs by text message instead of talking to a speaker. It got built and then removed the same evening, once it turned out the only real option was a business account and a compliance review for a service with exactly one user.