I don't use AI as autocomplete: I design systems with it. My main workflow runs on Claude Code — specialized subagents, MCP servers and custom commands — and I treat agent configuration as what it is: engineering, with rules, edge cases and maintenance.
The best example: an automated PR reviewer for Notable Learning. It validated that each branch matched a real Linear ticket, loaded the project conventions from a context repo, invoked subagents depending on what the diff touched (security, tests, architecture) and reviewed in two stages: first it checked whether previous blocking comments were actually resolved, citing file and line; then it did its own review against project criteria — multi-tenant isolation, roles as enums, Zod validation, zero PII in logs. All in a single consolidated comment, with severities and suggested fixes.
I designed it and ran it on my own PRs; the team reported that human reviews now arrived with the obvious issues already flagged.
I learned more from its failures than its successes: a noisy automated reviewer is worse than none, so half the design is the conditions for staying silent.
