Can AI Skills learn judgment not just rules? 👀
AI Skills are popping up everywhere. Salesforce in Claude recently launched with 37 sales Skills, and ChatGPT now supports Skills too.
Most design Skills I’ve seen focus on things like tokens, components, and repeatable checks.
I wanted to try something harder: research judgment.
So I built a Skill to help with usability research synthesis.
I gave it NN/g’s severity scale, then added a safety rule of my own: if something could seriously impact a user, don’t lower its priority just because it showed up once.
Then I tested the same research notes with the Skill off vs. on.
Without it, Claude still found the biggest risk. But it also mixed up some evidence, pulled a date from the wrong finding, and sometimes trusted what users said over what they did.
With the Skill, some mistakes still happened.
So I updated it: keep every detail tied to its source. If you don’t know, don’t guess.
I tested again. Better.
Then I checked every claim myself in a simple spreadsheet: claim in one column, source in the next.
I still caught a couple small slips. 😅
My takeaway: Skills are pretty powerful. 🤩 They gave the AI better guardrails, and each update made the output stronger. I still checked the work, but I’m excited by how much more useful the Skill made the process.
Curious: are you building Skills for rules, or for judgment? 👀
Join in the discussion on LinkedIn.