Testing
Testing not exhaustive but I've noticed the behaviour over months.
Skills added to ChatGPT exhibits the behavior listed below.
Grok without the Azure skills does not exhibit the behaviour.
Observed behaviour
Official Azure agent skills confuse AI models because they flood the context window with overlapping documentation, conflicting instructions from local repo files, and a mix of knowledge layers versus execution layers.
Instruction Collisions: Extensions inject tool descriptions and skill definitions that frequently fight with local repo instruction files (like .github/copilot-instructions.md or AGENTS.md) already loaded in the context.
Layer Mismatch: Azure Agent Skills act as a massive static knowledge layer (193+ cheat sheets), while actual execution layers like Model Context Protocol (MCP) servers handle active deployment, causing the agent to trip over whether it should just talk about a procedure or execute it.
Context Bloat: Preloading too many natural-language skill descriptions causes progressive disclosure to misfire, triggering irrelevant routines or diluting the model's focus.
Testing
Testing not exhaustive but I've noticed the behaviour over months.
Skills added to ChatGPT exhibits the behavior listed below.
Grok without the Azure skills does not exhibit the behaviour.
Observed behaviour
Official Azure agent skills confuse AI models because they flood the context window with overlapping documentation, conflicting instructions from local repo files, and a mix of knowledge layers versus execution layers.
Instruction Collisions: Extensions inject tool descriptions and skill definitions that frequently fight with local repo instruction files (like .github/copilot-instructions.md or AGENTS.md) already loaded in the context.
Layer Mismatch: Azure Agent Skills act as a massive static knowledge layer (193+ cheat sheets), while actual execution layers like Model Context Protocol (MCP) servers handle active deployment, causing the agent to trip over whether it should just talk about a procedure or execute it.
Context Bloat: Preloading too many natural-language skill descriptions causes progressive disclosure to misfire, triggering irrelevant routines or diluting the model's focus.