Map dependencies, not heroes
When every difficult question goes to one engineer, the immediate result may look efficient. The longer-term question is whether the team can make sound decisions when that person is unavailable. Treat concentrated knowledge as a system dependency, not as a personal failing or a reason to undervalue expertise.
List the activities that currently need a particular person: diagnosing a recurring incident, approving an interface change, restoring a service, or explaining an undocumented business rule. Distinguish rare specialist judgment from routine work that has simply never been taught. They require different responses.
Ask the specialist where interruptions are most costly and where mistakes would have the largest consequences. Prioritize a small number of high-value transfers. Trying to document everything at once can create a second workload without reducing the dependency that matters.
Transfer decisions through practice
Pair on a real, bounded task, but make the learner drive the investigation. The experienced engineer should explain how evidence changes the next step: which signal rules out a hypothesis, which assumption needs checking, and when to stop and escalate. A recording of commands alone misses that reasoning.
Use a controlled simulation for incident practice. Supply synthetic symptoms and an isolated environment, then ask the learner to identify the likely fault and propose a recovery plan. Do not create a live outage to manufacture a learning opportunity.
Reverse the roles on a second exercise. The learner explains the diagnosis while the specialist observes. Success means demonstrating the agreed task and recognizing its limits, not memorizing a sequence or becoming an instant replacement for years of domain knowledge.
Write runbooks with stop conditions
A useful runbook starts with a recognizable symptom and ends with a way to verify the outcome. Include prerequisites, required permissions, read-only diagnostic checks, escalation contacts, and conditions that make the procedure inappropriate. Link to the authoritative source instead of copying credentials or sensitive data.
Have someone other than the author follow the diagnostic path in a test environment. Note where they need unstated context, where names no longer match, and where a result could be misinterpreted. Those moments are concrete editing tasks, not evidence that the reader should have known better.
Shared knowledge does not mean shared unrestricted access. Preserve approval boundaries and least-privilege permissions. For industrial or safety-related systems, training material must not authorize control changes or bypass qualified human review. Knowing when not to act is part of operational competence.
