How to ask
The assistant can see your document. It cannot see what you meant. Almost every disappointing answer comes down to one of four things — scope, units, verification, or asking it to guess.
Name the scope
“The model”, “this view” and “what I have selected” are three different jobs with three different blast radii. Say which one you mean and the answer stops being a negotiation.
Without a scope this could mean four walls or four thousand. Destructive tools deliberately refuse a whole-model default and ask you to opt in, so being vague costs you a round trip anyway.
Renumbering has to know an order and a starting point. Supplying both turns a clarifying question into a result.
Say the units
A bare number is ambiguous in a way that is invisible until it is wrong. Write the unit — 1200 mm, 0.5 m, 4 ft — and it converts rather than assumes.
Internally these APIs work in feet. The assistant converts for you, but only once it knows what you started from.
Let it look before it acts
For anything that touches more than a handful of elements, ask for the list first. You get a chance to spot the wrong filter before it is applied, and the assistant keeps the result to work from.
Two steps, one decision point. It costs one extra message and removes the whole class of “that is not what I meant” outcomes.
Ask it to report back, not to reassure you
A good answer says what changed and what did not. If it tells you “done” and nothing else, ask for the breakdown — the tools return per-element success and failure, and there is no reason to accept a summary that hides the failures.
The second question is answerable from what the tool actually returned. The first invites a guess.
Undo works — use it as your safety net
Every write runs inside a single named transaction, which means it lands in the host application’s own undo stack as one step. If a bulk edit is not what you wanted, press Ctrl+Z in Revit, AutoCAD & Civil 3D or Navisworks exactly as you would for anything else.
Give it a target, not a puzzle
When intent is ambiguous the assistant is built to pick a sensible default and get on with it rather than interrogate you. That is usually what you want — but it means a genuinely ambiguous request gets a confident answer to the wrong question.
“Clean up” has a dozen readings. Naming the first move makes the rest of the conversation productive instead of speculative.
When there is no tool for it
The named tools are not a ceiling. If what you want falls outside them, say so — the assistant writes C# against the live document and runs it there and then. A bespoke aggregation, a one-off bulk edit, a report shaped to your office standard.
“There’s no tool for this: write code that lists every family type used on fewer than three instances, with its category.”
Submitted code is screened before it runs, against a blocklist covering the file system, process launching, network access and the registry. If your deployment needs a fixed tool surface, the whole feature can be switched off and everything else keeps working.
Write in your own language
Ask in Turkish, German, Spanish or English — the assistant understands the request and knows the local vocabulary for the products’ own concepts. Its replies follow the language you wrote in. The tool names and the parameters they touch stay in English, because that is what they are called in the applications themselves.
Next
Worked examples for your product: Revit, AutoCAD & Civil 3D or Navisworks.