copilot-instructions.md, AGENTS.md, SKILL.md, prompt files — they all tell the agent how to behave. So why does one get followed immediately while another disappears into the void? A practical mental model for how AI instruction files are actually loaded and applied — and how to spot guidance that is being ignored versus never loaded.

copilot-instructions.md, AGENTS.md, SKILL.md, custom agent definitions, prompt files — most repos that lean into AI tooling already have a few of these, and they all seem to do the same thing: tell the agent how to behave. So why does one get followed immediately while another seems to disappear into the void?

The answer is not in the markdown itself, but in when and how each instruction type gets loaded. Some are always in the room, some only appear when the task matches, and some do nothing at all unless you invoke them deliberately. In this session, Hidde breaks down those differences through practical examples and shows how small misunderstandings turn into silent failure, duplicated guidance, and unnecessary context burn.

Rather than handing out a giant compatibility matrix, you get a simpler way to reason about instruction scope, trigger conditions, and intent. The goal: the next time someone wants to add “just one more instruction file,” you’ll know whether it belongs there at all.

You will leave with:

a clearer mental model for how AI instruction files are actually applied
a way to spot when guidance is being ignored versus never loaded
better judgment about which instruction mechanism to use for a given job