Look for maintenance loops

I've experimented with a couple of different loops recently, both at Pactum and outside. For now, there is one I reliably get value from: a simple weekly hygiene loop for my website.

At its core, the loop is a set of written instructions following the Blomfield loop format, stored as a Markdown file in the site’s repository. Once a week, the loop gathers data, analyzes it, takes action on issues it finds, saves its reasoning trace, and shuts down. The full loop definition is below, in about 250 words.

In past executions, it has done useful work.

Created backups after the automated backups failed due to GitHub Actions billing limits.
Fixed a typo.
Fixed various issues with querying Ghost and Google Analytics data.
Closed stale issues.

But it's not been groundbreaking. I've also tried higher-value loops in this blog context: generating ideas for what to write about, or which people in my network to reach out to for coffee. But I never used the ideas they produced. The mistake I made there was asking for recommendations, which just creates more work for me. I need to put loops on tasks they can finish without handing the remaining 90% of the effort back to me.

So far in my experiments it feels like fully autonomous loops are a good fit for boring hygiene and maintenance, not leaps of innovation or even small, meaningful steps in a new direction. It's a bit like self-driving luggage: I lead in a direction, and AI pulls all the weight and stays behind me.

Could you just call it an agent? Perhaps, though everything is an agent these days, so the more specific term is useful. And even in the narrower sense, agents are often triggered by an external event and don't require as much prescribed structure.

You could also say it's just a scheduled prompt. But it has a structure, most importantly the compounding effect of saving its traces so future sessions can learn from past ones.

Weekly loop

Refer to loops/README.md for general instructions about loops.

Frequency

Weekly. "Window" refers to the time period between two consecutive loop executions.

Purpose

Keep taivo.ai operating smoothly, bug-free, as a modern content website.

Sense

Use the previous trace's completion timestamp, not its calendar date, as the window boundary.

Previous loop execution traces
Open GitHub issues and PRs, or ones closed or commented on during the window.
Analytics you have access to, per this repo
Live site for the posts published in the window, including email newsletters if applicable.
My notes, especially daily notes, that reflect what I am thinking about, though not completely.

Decide

Do not reopen or duplicate an issue before an explicitly recorded revisit date unless the impact materially changes.

You are allowed to

Fix straightforward bugs in content (e.g. grammar errors) or code (e.g. tracking bugs).
Propose larger fixes or improvements for the user's approval.
Revert or modify changes made by previous loop executions.

Act

For fixes you are allowed to make, push straight to master or edit Ghost.
For things requiring user approval, make a pull request.
When finishing execution, write your findings and reasoning briefly into the loop execution traces folder. Keep references, start with a 200-word executive summary in Simplified Technical English (STE). Include the model name and version.

Verify

Before putting any changes live, verify them (including visually) on the local Ghost install. If available, run local tests.

Learn

Propose what tools or loops might make this repository and loop more effective in the future. Act on these in accordance with the rules above in the "Act" section.