How to keep your AI moving when the data says no.
I had almost finished this post when I came across this video. I watched it, and I had to go back and rewrite parts of what I’d already written.
That doesn’t happen often. But the video clarified a few things I had touched on but not fully unpacked. And honestly, I’m not sure I would have understood the core message if I hadn’t written the first two posts in this series first. Some things only click once you’ve built the context yourself.
There’s a lot in the video, but three things stood out for me:
The handshake.
In my earlier posts I called it “the handshake” — the name and description of your SKILL.md file. The video makes clear this isn’t cosmetic. It’s the only thing your agent reads before deciding whether to load a skill at all. If it’s vague or wrong, the skill never triggers. Get this right before anything else.
Reference files are not attachments.
I was treating them a bit like supporting documents. The video reframes them as the place where all your context, rules, and knowledge live — separate from the process. The SKILL.md describes what to do. The reference files describe what to know. Keeping these two things apart…
