This site is a place for me to write. But now that it is working well I’ve found that I spend more time than I probably should finding little things for the coding agents to chew on. I enjoy watching them work, second-guessing their choices, discussing my wishes with them as they toil away.

Recent prompts (conversation starters, really) have included

I’m a little bit unhappy that the security audit script output doesn’t resemble the output of other test scripts. Those aren’t exactly common either. What do you suggest?

Review this codebase for security issues, custom code that doesn’t need to be custom, unused custom code, and other bloat. Don’t change anything, just fully understand the codebase and make a review report covering the four areas mentioned above.

There’s something awkward about how we’re generating changelog entries – the entries are generated on deploy based on commit messages, which means we need to commit again after committing and deploying. Commit, deploy, commit. Is there a better way to do this?

I occasionally have powerpoint or google slides presentations that I’d like to add to my portfolio. These often have speaker notes, and my typical approach is to put these speaker notes as a figcaption on the figure that represents each slide (and contains an image of each slide). To do this I have been exporting a PDF, running a script, then painstakingly copying and pasting in the speaker notes one-by-one. As you can imagine, this takes a long time, is error prone, and hard to do on a laptop screen. I’m considering a script to automate this entire process, so that all I would have to do is edit the speaker notes to be appropriate for a web visitor to read. Can you talk me through different ways we might accomplish this? (I’ll need to preserve the current PDF-to-image-to-figures script for a while.)

https://modern-css.com is very interesting. Are there things there that this project might learn? I’m thinking especially of things like selector nesting, aligning nested grids, balanced headlines, vivid colors, preventing layout shift, dark mode colors without duplicating values

But by pursuing these I’m polishing invisible parts of the site – the scripts I use to generate and test it, the structure of the CSS behind the scenes with limited visible effect on the product itself, the authoring process rather than the authoring results. It’s paint on the back of the cabinet. I like that it’s there; you will likely never see it.

The thing I will NOT do is have an agent write, or even start, a post for me. There are people at work who have an LLM write their documents, and, honestly, that’s fine. I review those no differently than if they had been completely hand-crafted. The person needs to stand behind their work and, since it is part of their job they should fully understand it; the usual difficulties of vagueness, flabby writing, jargon, and solutions masquerading as problems appear whether or not an AI model is involved. My review, shrugging off the “do not read me” sheen that every LLM-penned document seems to have on it these days, demands that the work meet its goal regardless of the level of their initial effort.

But for me, getting my thoughts out where I can see them means I must face the blank page.