The truth is that LLMs are mostly unpredictable and, as such, you can’t rely on AI agents to strictly follow the instructions stored in a markdown file. Even when they execute all the instructions contained in the skills, they often leave a backtrail full of trash/dirt. Because they’re often unruly, they need another chatbot to put them in line, as well as a human in the loop, of course. Basically, it’s a lot of trial and error…

“I acknowledge that “programming an LLM” is putting it optimistically, as skills aren’t usually deterministic scripts. But I like to think of them this way, and keep refining the skill until it yields the consistent result that I want.
Overall, I’m persuaded that tech writers who can build successful skills to automate their tasks will be on their way to the 10x tech writer goal (if that’s your aim). The best way tech writers can free up their time is by creating skills to attack those repeatable tasks (like release notes) since repeatable tasks keep chipping away at our productivity week after week. If you can fashion a skill that handles those recurring tasks, then you free up a recurring amount of bandwidth each week.

Additionally, most repeatable tasks fall into the category of mechanical toil that we want to automate with AI anyway. If we can automate the repeatable tasks, then we’ll have more time to tackle the one-off complex tasks that don’t fall into our laps weekly or biweekly.”

https://idratherbewriting.com/blog/all-about-skills-intro

#AI #LLMs #AIAgents #AgenticAI #Skills #TechnicalWriting #SoftwareDocumentation