Stop Repeating Yourself: Google Engineer Turns Agent Constraints into a Reusable Skill Library
Every repeated instruction you give an AI agent is a failure of memory, not of intelligence — this skill library makes those constraints persistent and checkable, cutting both rework and the manual diff-review burden. Teams running AI assistants in daily development get a concrete way to stop re-explaining the same rules in every new session.
The recurring pain with AI coding agents isn't that they can't write code — it's that every engineering constraint you explain (read the existing code first, only touch the named files, run lint before commit) lives only in that one conversation and evaporates when the session ends. The agent re-derives your working style from scratch each time, then "fixes" three extra classes and leaves you to audit the diff by hand.
addyosmani/agent-skills, which hit GitHub Trending on 2026-11-12 with 812 stars that day, takes a different angle than switching to a stronger model or writing longer prompts. Instead of disposable prompt text, it standardizes everyday engineering practices into a library of skills the agent can read and check against persistently. The author argues this reduces rework more than prompt tuning does, and the project is aimed at teams who are already using AI assistants daily but are burned by "changes too much, afraid to merge."
The article also offers three no-install steps you can take today: write rules into an AGENTS.md, make the agent report its intended file changes before acting, and archive any instruction you've said more than twice. It's a process-layer fix, not a model-capability upgrade.
Prompts are inherently disposable; encoding constraints as persistent skills treats the real bottleneck as context loss, not model competence.
The three suggested steps show that most agent rework can be prevented with lightweight process changes, not heavier tooling or bigger models.
The article deliberately distinguishes "the direction is worth following" from "the project is good to use" — a useful brake against treating Trending buzz as validation.
The rule 'say the same thing a third time, archive it' is the most practical. There's an extra layer of problem in a team: you archive it into your own config, your colleagues' agents still don't know, and when you hand off to someone else you have to repeat it all over again. Later we kept only one copy of rules and skills, put them in the shared environment everyone uses, with two or three people responsible for changes, and the rest just bring them along when assigning tasks.