To be expected, given how many organisations now require employees to use AI if they want to meet their OKRs, especially all that sell AI tools.

What's dumb, on top of everything, is needing to store non special standard operating procedures in specific AI folders and files when wanting to work with AI tooling.

Copilot today supports the top-level AGENTS.md approach as well, which seems to be the cross-tool "standard".

It is a standard in a sense that they will all read it (although last I checked you still need to adjust the default config with Gemini). But feature support varies between different tooling. For example, only Claude supports @including other files.

The "standard" AGENTS.md suggestion for that is [regular markdown links](./like-this.md)

The problem is that it doesn't actually include the referenced file in the context. The model will only see what's in it if it deigns to read it, but that's not a given in all circumstances where it might need to.

I use this feature often in Claude to bring specific files so that they are in context at all times. E.g. when working on a parser, I will often put the grammar to be always in context. Or if working on a web app, all the model types.

Like needing to store IDE specific files?

Both can be true.

> Extremely simple changes do not require explicit unit tests.

I haven't used Copilot much, because people keep saying how bad it is, but generally if you add escape hatches like this without hard requirements of when the LLM can take them, they won't follow that rule in a intuitive way most of the time.

It is kind of alright, I use mostly on VS when coding C# or C++, for code completions, error analysis, check code quality and such.

As agent, or writing everything for me, not yet.

the $10 plan makes a great backup to claude or codex and the inline completions are nice

Yeah, I tried various very sane-looking instrucions file when starting to use copilot 6 months ago. Turned out it was not really useful. It mostly follows the rules anyway, but it also often forgot to. So turns out, especially with the fast turnaround with models today, it was better to just forego these instructions files.

It doesn't say much really. At this point we can assume almost every project has some generated code in it. Unless you're sure that every single author hates the idea and there are no external contributions. Agent configuration just makes it clear.