Help
How to ask so the assistant finds it
Naming things beats retrying. A few habits that make answers land on the first try.
Updated July 17, 2026
How to ask so the assistant finds it
You never have to phrase things carefully — asking naturally works most of the time, and when it doesn't, that's ours to fix. But if an answer misses, one rephrase almost always beats five retries. Here's the shape of it.
Name things
Every pronoun is a small puzzle the assistant solves before it can search. "My", "it", "that repo" — each one is an inference step, and inference steps are where smaller, faster models fumble. Naming things removes the puzzle entirely.
| Instead of | Try |
|---|---|
| "What were my commits yesterday?" | "Commits by draftin1 in the dialoguesai org on July 15" |
| "Did she reply about it?" | "Replies from Priya about the renewal quote" |
| "What's on my calendar this week?" | "Calendar events July 14–18" |
| "Show me that doc again" | "The doc called 'Q3 planning notes' from Drive" |
The pattern: accounts by name, places by name, dates as dates. You don't need all three every time — add the one the assistant got wrong.
Retry once, differently
If an answer misses, don't repeat the same question louder — change the part that failed. Check the answer's receipt (the reasoning trace) to see what it searched for, fix that one thing, and ask again. Reading the receipt takes ten seconds.
Harder question, stronger model
Model choices trade speed for care. The fast options are great for everyday questions; a question that needs many steps in a row — search, cross-reference, summarize — is worth switching to a stronger model for. You can change models right from the chat. If a multi-step answer came back half-done, the switch is usually worth more than a rephrase.