Describe outcomes, not UI flows
Avoid "open the mesh dialog and click X." Say what the simulation must demonstrate. The agent owns the how. A good prompt names the physics, geometry source, QoIs, validation criteria, and the credibility bar ("within the cited test-rig uncertainty", "safety factor >= 2 per ASME BPV VIII-2").
Review the typed plan before approving
When the plan card appears, scan the CaseSpec changes the agent intends to make. If you disagree with a boundary condition, a turbulence model, or a mesh target — say so. The agent will patch the typed state, not rewrite the whole plan from scratch.
Ask for evidence, not opinions
If the agent makes a non-obvious claim — "y+ is within wall-function range," "the residuals show convergence" — ask which artifact backs it. The agent will link to the parsed monitor, log line, or function-object output. If it cannot, the claim is wrong.
Load skills when stuck
Skills are signed workflow guides. Type a slash command (or click a help-drawer suggestion) to load one. The relevant skill text becomes part of the agent's context for the rest of the turn. Useful when migrating between solvers, adopting a new standards pack, or onboarding a new physics regime.
Interrupt cleanly
You can interrupt a running tool call. The agent records the interrupt as a typed event, captures what was already produced, and reverts to plan mode. Restart the next attempt from the typed state, not from "what did we just do."