5 What I Learnt This Week

I run AI directory sites and a YouTube automation pipeline. This week, I hit several points of friction. Here are five lessons from my adjustments.

  1. Control Your CI Costs

My Bluesky posting script was wasting GitHub Actions minutes. Every post triggered a massive build process across three sites. I was spending 120 minutes a week on simple status updates.

I made two changes:

  • I switched from three daily triggers to one single daily trigger.
  • I added a path filter so text edits do not trigger a full site rebuild.

Do not let small tasks burn your automation quota. Fix these habits before your project grows.

  1. Add Quality Control to Automation

I found 17 posts in my queue that sounded like a bot. They used phrases like "auto-generated" which felt wrong for my personal brand.

I added a QC gate to my pipeline. This step checks posts for:

  • Broken links.
  • Expired news.
  • Robotic or spammy tones.

If a post fails, it stays in the queue for manual review. I post less often now, but the quality is higher.

  1. Simplicity Could Beat Optimization

I tried removing AI model routing. I used to send simple tasks to cheap models and hard tasks to expensive ones.

After removing the router, I found:

  • Latency stayed the same.
  • Costs rose by 8%.
  • The code became much simpler.

The 8% cost increase is worth it to avoid debugging routing errors. At a small scale, complexity costs more than the API savings.

  1. Watch Your Licensing

I added image slides to my YouTube tool using Openverse. The default results include many Creative Commons types.

If you do not filter for CC0 or PDM licenses, you might use images that require on-screen credit. For a monetized channel, this is a legal risk. Always filter your API requests upstream to avoid accidental copyright issues.

  1. Monitoring Tools Depend on Usereasiy

I tested Netdata, SigNoz, and OpenObserve.

  • Netdata is easy and works immediately.
  • SigNoz needs you to instrument your code with OpenTelemetry.
  • OpenObserve is great for logs but has a steep learning curve.

For my current setup, these tools felt like overkill. I chose a simple error alerting integration instead. Pick the tool that fits your current infrastructure, not the most complex one.

Source: https://dev.to/morinaga/5-things-i-noticed-this-week-ci-cost-bluesky-qc-and-cc0-licensing-49ig