The chat metaphor held me back for too long. These are black box functions that produce text and should be treated as such.

It came to me gradually that I was slowly replacing each part of my skills with scripts that took defined inputs, validated them, triggered agent sessions with defined prompts and spit out validated outputs in defined formats.

Now I do it proactively, if it looks like a function or a loop, or workflow encode it as such. My outcomes are better, I can use cheaper models and it’s easier to test.

▲

kingkongjaffa 15 hours ago | parent [-]

Do you have an example?

	▲	skydhash 14 hours ago \| parent [-]
		Not GP and I’m a general AI opponent, but it is the same strategy I adopted at work due to mandates. I can’t give you a past example(they’re all related to $work). But you can probably work out a skill that downloads an emacs package, extract all user options, provide general configuration frameworks, prompt you with those and then write the use-package incantation for that. The pattern is to systemize your decision making, figure out decision node, and then provide a framework to solidify that structure. It will rigidify it, but you gain in speed that way. Basically writing like the AI is an idiot. It’s not, and it’s not intelligent either. Bit you have to make the process explicit. In most cases, a deterministic script would be better. The AI advantage is flexibility. In the above example, the flexibility is required for the configuration recommendation. Everything else before that component, and after should be very explicitly and even specified as small software to make it stricter and less prone to errors.