People often find themselves confused when they encounter financial service agents released by Claude Code, only to discover they’re just a collection of skill.md files. Are agents really that simple? Aren’t they supposed to be complex codebases filled with numerous coding files?

Well, it could be both. So, when I was attempting to create a custom factsheet generator agent, I explored both approaches:

• SKILL.md approach → Good for ANALYSIS (reasoning, adapting, exploring)
• Python codebase → Good for PRODUCTION (recurring, reliable, fast)

First, let’s see where and why Skill.md agent excels

But if you use codebase agent, it can be quite specific, rigid/accurate, repeatable

Recommended middle ground, we could build both, keep thin:

The SKILL.md describes what the skill does and how to invoke it, but delegates execution to the Python pipeline. You get:

• Plugin ecosystem compatibility (invoke via /factsheet-generator)
• Production reliability (Python handles the work)
• Easy monthly runs (one command, deterministic output)
• Maintainability (fix bugs in Python, not by rewriting skill instructions)