Capstone: Build & test

Build the way the course did — in versions, each one live-able: v1 deterministic rails (trigger, filters, formatters, outputs; test with real records), v2 AI steps behind validation (contract prompts, run the full test suite, wire the unsure lane), v3 gates and reliability (approval flow, problem folder, alert tiers, the break-it session). Resist building v3 in one sitting: each version catches its own class of bugs while the workflow is still simple enough to see through.

The testing bar (all four, no skipping)

• The suite: 12+ fixed cases with expected outputs, scored in a sheet — normal cases, each category/branch, empty-field cases, the ambiguous case that should land unsure, one injection attempt. No more than 2 failures across your suite before going live; every prompt edit reruns the whole suite.
• The break-it session: rename a target, force a timeout, reject an approval, feed garbage. Every failure lands where the pre-mortem said, or the pre-mortem gets fixed to match reality (also a pass — updated docs are the point of docs).
• The retry-safety check: for each step that sends or writes, answer the double-send question in writing. Any 'unsure' becomes 'fail to problem folder'.
• The five-second gate test: judge one of your own approval messages cold, the morning after building it. If you need to open another tab, fix the message before an actual reviewer meets it.

The two-week soak

Then: turn it on and run it for two real weeks alongside daily health minutes (60 seconds: run count, problem folder, approval rate) and a small operator log — date, anything weird, anything fixed. The soak is where automations earn trust: week one surfaces the edge case your suite missed; week two proves the fix held. Keep score against the success criteria from your design doc — those three sentences are about to become your verdict. Mid-soak prompt edits are allowed and normal; log them (changelog!), rerun the suite, keep soaking.