---
title: "Comment by Ahmet Özel on \"AI Agent vs Skill: Which Layer Holds the Logic?\" | Hashnode"
description: "The scraping example makes the code sidecar's responsibility important. A strict JSON schema can ensure the expected fields exist while the extracted values sti"
url: "https://hashnode.com/posts/ai-agent-vs-skill-which-layer-holds-the-logic/6afb1bb1f4b2b394a3c4ba30/comment/6afb74174523ddb00ebb7c24"
---

[Discussion](https://hashnode.com/posts/ai-agent-vs-skill-which-layer-holds-the-logic/6afb1bb1f4b2b394a3c4ba30)

Comment [AI Agent vs Skill: Which Layer Holds the Logic?](https://hashnode.com/posts/ai-agent-vs-skill-which-layer-holds-the-logic/6afb1bb1f4b2b394a3c4ba30)

[Ahmet Özel](https://hashnode.com/@ahmetozel)

[4d ago](https://hashnode.com/posts/ai-agent-vs-skill-which-layer-holds-the-logic/6afb1bb1f4b2b394a3c4ba30/comment/6afb74174523ddb00ebb7c24)

The scraping example makes the code sidecar's responsibility important. A strict JSON schema can ensure the expected fields exist while the extracted values still come from the wrong page element. Packaging the instructions preserves the process, but the model-dependent parts still need behavioral checks.

I would version a few saved DOM fixtures with the skill, including missing fields and two similar elements on one page. Running those through the same extractor and output validation would show which reliability gains come from deterministic extraction code and which depend on the model following the packaged instructions.

[techpotions](https://hashnode.com/@techpotions)

[1d ago](https://hashnode.com/posts/ai-agent-vs-skill-which-layer-holds-the-logic/6afb1bb1f4b2b394a3c4ba30/comment/6afb74174523ddb00ebb7c24#reply-6aff348b2ba2db8fedf2763b)

Exactly, the schema proves the shape, not that the right element filled it, and that gap is where skills quietly leak. The versioned-fixture idea is the right fix: a golden set with missing fields and two lookalike elements splits “deterministic extraction” from “the model followed instructions this time” far better than any output check can. We pin DOM snapshots the extractor runs against on every change for exactly this reason, adding the two-similar-elements case to ours, thanks for that.
