Over the past year I’ve built a set of shared AI skills and review agents for a web development team. Some of it changed how we work. Some of it we quietly turned off. Here’s the honest version.
Where it pays off
- Scaffolding repetitive code from a known pattern, like a new page builder block or custom post type.
- A first-pass code review against written standards, before a person looks at it.
- QA crawls for broken links, missing alt text, console errors and leftover placeholder text.
- Turning messy notes and meeting transcripts into structured specs and checklists.
- Drafting migration and import scripts that a developer then reviews and runs.
Where it doesn’t
- Vague requirements. AI fills gaps confidently, which is worse than leaving them empty.
- Debugging an unfamiliar production system with no logs to look at.
- Anything security-sensitive that ships without a human review.
- Design taste. It can follow a system; it can’t tell you the system is wrong.
What made the difference
Three things. We wrote our standards down so the AI had something real to follow. We kept review agents read-only: they report, a person decides. And we built many small, repeatable skills instead of one giant prompt that tries to do everything.
AI multiplies your process. If the process is undocumented, it multiplies the chaos.
