Our SDK team leans on LLM tooling daily as part of its CI and release workflow. That platform is a team effort; below are the pieces I personally authored and own — the automation and guardrails that made the AI parts dependable and lowered day-to-day toil. I've scoped this deliberately to my own contributions.
Releasing across an ecosystem of 20+ ad-network adapters is repetitive and error-prone. I built a /release-adapter skill — a single-command driver that releases one adapter end to end — turning a fiddly manual checklist into a repeatable, one-line operation.
Authoring test content by hand doesn't keep up with the SDK's surface. I worked on the headless LLM authoring path in the automation pipeline — generating test content programmatically — and made it robust in CI, including surfacing the model's error clearly when headless authoring fails instead of letting the run die silently.
Generic AI feedback isn't useful on a large, opinionated codebase. I added Claude Code project-context documentation to the monorepo — architecture, conventions, and the public API contract — so any AI tooling operating on the repo is grounded in our standards rather than defaults.
The team runs an automated review step in CI. I built the ergonomics and safety valves around it so it stays useful and never becomes a hostage situation:
/skip-ai-review — a deliberate, audited bypass for when a human overrides the model./re-review — re-run cleanly after pushing fixes; I also fixed the case where /skip-ai-review and /re-review could cancel each other out.I explore the same ideas from scratch, and fully solo, in my C++ engine's agent harness.