Blog Post
Skill or Sub-Agent. Choosing AI Capabilities You Will Actually Reuse
EvdV You are right on the mechanics. A skill is text. The model matches it, loads it via a tool call, and from there it is basic inference. No hidden runtime and banned_phrases or self_check are illustrative rather than reserved keywords.
Where I would push back is on dismissing the concept itself. An executable contract is an architectural pattern you author, not an out-of-the-box model feature.
In production that looks like this. Following the core skill, the agent runs a deterministic lint tool or evaluator. It scans the draft against your configured rules and blocks the output until it meets the contract. From the outside it feels like the skill corrects itself. Inside, a bounded tool does the deterministic work.
That is the article theme in miniature, separating the user experience from the engineering shape underneath. A useful distinction to draw out. Happy to share a pattern diagram of how we wire these validation passes if that helps.
yeah but that was the point I wanted to make - anything you want to enforce you need to build yourself, it doesn't happen magically -
- KishoreKumarPattabiramanAug 12, 2026
Microsoft
Exactly. That is the boundary between what the base model gives you and the architecture you author on top. Glad we are on the same page.