battle-testing-on-real-infralisted
Install: claude install-skill pedro-angel/agent-methodology
# Done Means Validated Live, End-to-End
Nothing involving an external system, a deployment, or a durability/resume guarantee is "done" until it has run against real infrastructure — genuine instances provisioned for the test, never the production estate — end-to-end, with the result captured as evidence. Mocks prove your wiring; only a live run proves reality.
## When to use
Reach for this skill whenever you are about to mark complete: an integration with a third-party API or provider, a deployment to a real environment, or any hard guarantee (durability, idempotency, cross-process resume, exactly-once, failover).
Red-flag thoughts — if you catch yourself thinking any of these, STOP and apply this skill:
- "The mocks are green, so it works."
- "The design guarantees it survives a restart / the docs say the endpoint returns X."
- "It's basically the same as the path we already tested."
- "I'll write the live test later / it's too slow to run every time."
- "The SDK handles that for us." (Did you watch it handle it?)
- "The evidence is more credible with the real hostname in it." (Properties prove; identity only locates you.)
- "CI is green, so it works." (Does CI actually run the live suite — or is it wired into no gate at all?)
## The rule
1. Re-assert "done" against live infra. Provision the real services (a real database, a real deployment target, a real provider account — instances stood up for this test, never the production estate), run the real agent or code path