I used Claude Code to build Agent Skills. A planner skill, an implementer skill, ... Took a few iterations to get the basic. They worked. Sometimes, I revisisted to improve them.
The three questions for Agent:
The three questions for Agent:
- What can we improve?
- Which parts are not good?
- Which parts are not useful?
The first question always generates a lot of stuff. It seems more is "improvement." I got something from Agent.
The second question forces Agent to analyze and evaluate its work. Amazingly, it pointed out many things. Again, I got some improvements.
The last question put Agent in different mode. "Is this sentence/part useful?" It trimmed lots of stuff, filler words, sentences, nice to have stuff but no actual value.