I had a pretty naive idea on day 16: take my bookmarked posts on X, ask AI for 1-2 hour experiments and get a full backlog. Ideally this would be built in a few hours of prompting. The hiccups started with X login/session problems. Then my browser automation was getting flagged for suspicious activity. I had to let things cool down for a few days. Today (day 24) is the day I continue.

First I continued on the road I was already on by having my X posts scraped automatically in Chrome with Playwright. This kept having me have problems, so I thought what if I just ask Codex to use my Safari browser instead. Safari already had an active X session that seemed to work fine. Could it be that easy and just change the browser automation? After changing some code in the app it just worked with Safari instead. Something I didn’t expect to be so easy. I wasn’t fighting the platform anymore. I was already logged in here as myself, a human, so not a bot.

After the first successful run I checked my OpenAI API usage to make sure I am not spending a shitload of money on these experiments. After just 1 created AI experiment it racked up a $0.23 cost for a simple experiment text. With way more experiments needed and multiple runs it would quickly add up seriously. So I changed the model to the gpt-5-nano-2025-08-07 model instead. The costs were down dramatically per call, without meaningfully less quality. In the past, the costs alone would have been enough reason to abandon the project.

# AI Experiment Backlog
Last generated: 2026-03-13T10:48:02+00:00
Bookmarks processed this run: 4
Cards added: 3, cards updated: 0

Every compromise in this project, like the browser switch and the cheaper model, was a small act of choosing “done” over “perfect”.

These daily AI experiments force me to ship and write about it, every single day. A huge thing for someone struggling with perfectionism. As this daily cadence makes sure my perfectionism suffers from it. Great, just what I needed. I am starting to enjoy shipping and writing about (half) completed and short iterative cycles.

Every time I tried to make the system cleaner or more automated, it broke. The version that eventually worked was the one that cooperated with the mess: my existing browser, a cheap model and incremental runs. The final product was not “AI reads my bookmarks.” It was “I built a cheap, reliable habit collector out of messy inputs.”

Key Insight

I used to think that I was abandoning projects because of technical problems or other hiccups. Writing this post made me see that my perfectionism (again) was the culprit. Seeing this project semi-finished, full of compromises, and still getting a write-up is a good sign for me.