Enjoyed the course. Good to learn how MCP, skills and agents are related.
For the sales analysis example - why is AI needed at all. Why not just process the sales data as a batch script - much more reliable. Then just use genAI to create the powerpoint or summary doc at the end?
Ditto with the time series data example - the ‘skill’ is defining lots of python code. Why not just invoke all that directly rather than hope Claude figures it out. Is the idea that non-programmers write these skill files so that coders aren’t needed ?
I marvel at Claude’s ability to plan and execute skills, but for these deterministic workflows wouldn’t regular sequential code be more reliable?
Your point regarding why not use deterministic code is valid. Not only is the code deterministic but also executes faster and at a lower cost.
Skills continue the movement from punch cards, assembly language, scripting all the way to English level programming. A big adoption step for many. This is a very big step forward. Though using Python in a Skill may be difficult to debug.
Given how easily Claude Code has made programming it seems to make sense to use each for their strength. Deterministic code for its strength and skills for theirs.
Looping would need to be added to Skills before I would consider adoption. A beauty of today’s paradigm is that a model can evaluate unstructured and semantic content which used to be a program stopping point … “wait for human intervention”. Today, a model’s ability to understand context enables the program to continue processing the loop because it provided the feedback that a human used to.