Claude Code plugins that run every day on live business operations, not demos. Each one was built for a real workload first and generalised afterwards.
The common thread: a language model does the judgement work, and deterministic checks with exit codes decide what is allowed to leave the machine. Rules that only exist as prose get ignored eventually. Rules that are scripts do not.
claude-email-support-plugin Customer support that answers itself, with four blocking gates deciding what is safe to send. Live state diffing against payment and membership records, failures harvested into regression tests. 578 production runs.
claude-ig Full-lifecycle Instagram content engine. 13 sub-skills, 6 specialized agents, and a 5-category scoring model.
- Gates over instructions. Anything that touches a customer passes a script that returns a status, not a paragraph asking the model to be careful.
- Context derived, not configured. Prices, products and plans are read from the source systems on every run and diffed against the last known state. No list to maintain.
- Failures become tests. Every incident is harvested into a regression case, so the same class of mistake cannot come back unnoticed.
- Measured, then changed. Design decisions here come from reading production logs, not from assumptions about what should be slow or expensive.