A lot of AI advice for plants stays abstract. This post is the opposite. A Claude skill is just a repeatable task captured once so anyone on the team can run it again, the same way every time, on live data. The best first skills are the unglamorous recurring jobs your process and reliability engineers already do, the ones that eat hours and reward consistency.
Here are ten worth building first. Each is concrete, sits safely beside the control system rather than inside it, and keeps a human on the decision.
The starter library
Shift-handover summary. Read the last twelve hours of alarms, work orders, and operator log entries, and produce a clean handover brief the next shift can absorb in two minutes instead of twenty.
Work-order triage. Take the incoming work-order queue, group by unit and criticality, and draft a prioritized list the planner reviews and approves. The planner still decides; the sorting is done.
Bad-actor digest. Each week, rank the assets driving the most unplanned downtime, compared against their own baselines, so the reliability meeting starts from a current list rather than last quarter's.
MOC drafting assistant. Given a proposed change, assemble the first draft of the Management of Change package, pulling the relevant asset history and prior reviews. The MOC process still governs; the paperwork starts further along.
Inspection-report reader. Turn a stack of inspection PDFs into a structured, queryable summary, so "what did the last three inspections say about this vessel" is a question, not an afternoon.
Historian question-answerer. Let an engineer ask the historian in plain language, namely "show me where this exchanger's approach temperature drifted over the last month," without writing a tag query.
Procedure lookup. Wire Claude to the operating procedures so a control-room question returns the right step from the right document, with the source cited, instead of a hunt through a binder.
Spare-parts cross-check. When a work order calls for a part, check stock, lead time, and equivalent substitutes, and flag the gap before it stalls the job.
Daily reliability watch. A standing loop that reads yesterday's trends across rotating equipment and surfaces the few assets drifting toward an alarm limit, so attention goes to the right place each morning.
Incident write-up starter. After an event, assemble the timeline from alarms, logs, and work orders into a first-draft incident narrative the team edits, rather than reconstructing the sequence from memory.
Why these ten
Notice what they share. Every one of them is recurring, structured, and currently done by a skilled person spending time on assembly rather than judgment. Every one keeps the engineer on the decision and Claude on the legwork. And none of them touch the control loop; they sit on top of the DCS and historian, exactly where advisory intelligence belongs.
That's the test for an eleventh, twelfth, and twentieth skill too. If a task is repeated often, follows a pattern, and ends with a human decision, it's a skill worth building. If it's a one-off, or it requires a judgment call in the middle that only a person can make, leave it as a conversation rather than a skill.
How to start
Build one, this week, with the engineer who does that task today. Watch it run beside them for a few cycles, correct it where its judgment differs from theirs, and let them tell you whether it earns its place. A skill that survives that test becomes shared infrastructure the whole team reuses. Then build the next.
The fastest way to get a team fluent enough to build these well is hands-on. Our Practical AI Warmup is eight 1:1 hours that leave a manager or senior engineer with a working model of how skills like these get built and where to point them first.