The shift from typist to director moves the unit of skill in a way most developers haven't named yet. Code-craft is necessary but no longer sufficient. Two new disciplines do the load-bearing work now. Spec-craft: writing the framing, constraints, and acceptance criteria precisely enough that a competent agent produces the right thing on the first or second try, instead of negotiating across ten. Supervision-craft: reading agent output for the failure modes that don't trip your tests but will surface later. Both are real skills. Both decay if you don't practice them. The agentic engineers who get great are the ones who treat those two as their primary craft, not as overhead between the real work.
Lots to think about here. I use the tools for a couple tasks. I usually write a number of bullet points of what I want to cover in a post. Send it off to a custom skill to give me a first draft. I'll edit it then have another custom agent review it and ask me about issues or gaps it sees before it makes its own edits. I review/edit/approve. Publish skill sends it to where I publish. I do the "taste" parts, AI does the drudgery.
It's still vibe coding if the "Senior Director" is not an engineer. I catch subtle mistakes all the time that someone without a coding background would overlook within the agent output and recommendations. Some of these ill thought out plans, implemented at the early stage of a project, can have catastrophic effects when the application goes from development to production. Something developed for a handful of users can have dire financial consequences on token usage if put into production at scale. Just remember, agentic coding can save time and complete complex tasks, but you still need to check your work. I have also see agentic code that is so difficult to review, it would take more time to review, than to just start the project over done right. That kind of insight requires a human engineer!
The shift from typist to director moves the unit of skill in a way most developers haven't named yet. Code-craft is necessary but no longer sufficient. Two new disciplines do the load-bearing work now. Spec-craft: writing the framing, constraints, and acceptance criteria precisely enough that a competent agent produces the right thing on the first or second try, instead of negotiating across ten. Supervision-craft: reading agent output for the failure modes that don't trip your tests but will surface later. Both are real skills. Both decay if you don't practice them. The agentic engineers who get great are the ones who treat those two as their primary craft, not as overhead between the real work.
Lots to think about here. I use the tools for a couple tasks. I usually write a number of bullet points of what I want to cover in a post. Send it off to a custom skill to give me a first draft. I'll edit it then have another custom agent review it and ask me about issues or gaps it sees before it makes its own edits. I review/edit/approve. Publish skill sends it to where I publish. I do the "taste" parts, AI does the drudgery.
It's still vibe coding if the "Senior Director" is not an engineer. I catch subtle mistakes all the time that someone without a coding background would overlook within the agent output and recommendations. Some of these ill thought out plans, implemented at the early stage of a project, can have catastrophic effects when the application goes from development to production. Something developed for a handful of users can have dire financial consequences on token usage if put into production at scale. Just remember, agentic coding can save time and complete complex tasks, but you still need to check your work. I have also see agentic code that is so difficult to review, it would take more time to review, than to just start the project over done right. That kind of insight requires a human engineer!
https://myzservices.substack.com/p/short-video-of-why-you-should-not?r=70dhgg&utm_campaign=post&utm_medium=web