Criterias for instruction evaluation.

[RredCat]

Hi folks,

I am enjoying using instructions on my projects. However, I aim to assess the actual impact on GitHub Copilot.

I have read the topic - We need evals or something. But it will take time to implement, and frankly speaking, it will be too general (one project's trash is another project's treasure).

Do you have any criteria to evaluate the impact of instructions on a project?
I am tuning every instruction to the project's needs, but I don't know how to measure real impact.
I have instructed Copilot to respond in various human languages or adopt a Yoda-like tone. This impact is easy to spot.

Which criteria do you use to measure the impact of technical instructions?

[alpha912]

[RredCat]

Your answer looks like AI slop

[aaronpowell]

@alpha912 Please don't spam - consider this a warning for being disruptive.

[aaronpowell]

Measuring the impact is an inherently difficult task, as you note, so I can only really comment on it from my perspective.

When I'm using AI assistants for coding, I generally have an idea of what I'm trying to achieve, so I'm looking for a way to speed up the process. This is where AI comes in, I delegate one or more tasks to it, and when it's complete I can review it through the lens of "well this is what I would have done".

[alpha912]

Very interesting... It is such a crazy coincidence... I was using cursor cli with claude 4.5 to code something earlier... And now I was wondering why i got a strange github email notification.. And I see i commented on your discussion.. I don't have a clue what this repo or this discussion is.. It looks a lot like the agent used gh cli for some action and hallucinated and did something. I could not trace it.. But i guess we should really be more careful in delegation of coding to ai agents...