Writing a Good Skill is not a Simple Process

I mentioned earlier that Video Director development and testing is almost complete, and the last thing to do was to write a skill for Claude CoWork that makes video creation and editing simple and conversational.

It’s not as easy as I thought it would be.

Originally, I had Claude Code write the skill based on what I wanted. That seemed simple enough. The data model is not too difficult to understand:

• Every time you have new or changed functionality in an app, CoWork should know that the app has changed. There is a table for all the runs of recording videos of the app. All you have to do is look at the table and see if there is a material difference between the last run and the last push of the app into GitHub. That is not as simple as it might seem.

• There is a concept of a Journey for each app’s videos. An end to end recording of an app’s features is the first and default Journey. If I wanted to do an overview video (not an end to end), that is a different Journey.

• Within each Journey, there can be versions. I can have a version in English, another in Spanish, a third in French, etc.

It seems fairly simple to me, but I got a lot of pushback from CoWork saying the skill Code wrote did not align with what I was asking CoWork to do.I knew the data existed to make the skill work as desired, so I explained the data model. CoWork offered to amend the skill and proposed changes.

I told CoWork to write up the proposal so that I could run it by Code first. So that’s what I did - and guess what? Code says, “CoWork is right. I did not write this skill correctly. Let me fix it.”

Again, I had to have a little argument with Code. I know the data is anvailable, ans I designed the data model myself. Now I’m re-explaining the data model Code had already implemented.

That was a couple of hours of work, but still very much worth it. There is still a little more testing to do, as we inch closer and closer to a great AI skill.