Name the failure modes out loud
We list what can actually go wrong: data loss, privilege creep, a stuck cutover, a single person who holds the only working runbook. Naming them early is cheaper than discovering them in production. Each failure mode gets a severity guess and a first mitigation, even if the mitigation is simply "do not touch that yet."
Bound blast radius before you ship
A change that can take down every customer is a different animal from a change behind a feature flag for one cohort. We design for limited blast radius: staged rollouts, read-only first paths, dual-running periods, and clear kill switches. If blast radius cannot be bounded, the change waits until it can, or the business explicitly accepts the exposure.
Prefer reversible changes
Forward-only migrations and irreversible schema bets are common ways programmes paint themselves into a corner. Where we can, we choose steps that roll back cleanly: additive schema first, dual writes with a measured cutover, configuration that can flip back. Irreversible steps get extra scrutiny, a rehearsal, and a written go/no-go.
Keep secrets and privileges tight
Credentials in chat, shared admin accounts, and long-lived tokens are quiet risks that turn loud during incidents. We push secrets into proper stores, shorten lifetimes, and separate who can deploy from who can read production data when the context demands it. Least privilege is not theatre; it is how you stop one compromised laptop from owning the estate.
Plan the ugly path
Happy-path demos hide the work that matters under stress. For material changes we define degrade, stop, and rollback before go-live: who decides, which signals trigger it, and what "safe enough" looks like afterwards. If the ugly path is undefined, the team will invent one in the middle of an outage.
Write who acts when risk becomes real
Mitigations fail when nobody owns the next move. We attach a named owner and a contact path to the risks that matter, and keep that list next to the runbooks. Escalation should not depend on remembering who was in last week's meeting. Clear ownership turns risk registers into something people can use at 02:00.