Tooling a team kept using.

Most AI engineering stories are demos. This one is a plugin suite I wrote and maintained that my team at byrd used every working day, because it did the parts of the job everyone was doing inconsistently by hand.

Why it worked

Because each skill encodes a decision the team had already made and kept re-litigating. The value was never the model. It was writing down how we work, in a form the tooling could apply consistently.

The public version

I have rewritten it clean-room, with nothing employer-specific in it, and one significant design change. The original hard-coded my team's conventions: our branch format, our PR template, our migration style. That made it excellent for us and useless anywhere else, and actively harmful in a second repository, because it produced confident output in the wrong shape.

So the public skills discover conventions instead of carrying them. One skill reads the agent instructions, the contributor docs, the templates a repository already has, and failing all of that, the git history. Every other skill depends on it. When a convention genuinely cannot be found, the skill asks once and offers to write the answer down, so the next run does not have to ask again.

The six skills

• project-conventions. Works out how this repository does things. Every other skill calls it first.
• refine-ticket. Turns a vague ticket into something buildable, with scope, acceptance criteria and an explicit out-of-scope list. Decides what.
• plan-implementation. Reads the affected code and produces an ordered plan where every step names how it is verified. Decides how.
• open-pr. Fills the repository's own template, and refuses to invent content for a section it cannot answer.
• review-changes. Reviews against the standards a project states, separates defects from preferences, and constructs the failing case before reporting a defect.
• create-migration. Schema changes, in the project's own style, with the rollback actually tested rather than assumed.

No code and no dependencies: each skill is a single markdown file. That is the point. The artifact is the written-down decision, and the model is only the delivery mechanism.

Read them on GitHub. MIT licensed, clone it or copy the directories you want.