Best practices

Practical habits that make your assistant more useful and more predictable, drawn from how it's actually used.

Give context once, not every time

Your assistant remembers things you tell it — preferences, recurring context about people and projects, how you like tasks handled. Say it once explicitly ("always cc my manager on client emails") instead of repeating it in every task. See Memory for how to check what it's already retained and correct anything wrong.

Start narrow with supervision, then loosen it

When you're new to a skill or a type of task, keep it under closer review — let the dispatch cart hold actions for your approval before they go out. Once you've seen a few runs go the way you expect, widen autonomy for that specific kind of task rather than for everything at once. Blanket "trust everything" settings are where mistakes get noticed latest. See Supervision & visibility and Dispatch cart & execution.

Turn repeated tasks into workflows early

If you notice yourself asking for the same thing on a schedule — a weekly status digest, a daily inbox triage — don't keep asking manually. Save it as a workflow the first time you notice the pattern, not the tenth. It's easier to tune a workflow that's already running than to remember to set one up later.

Be specific about ambiguous instructions

"Handle my inbox" is broad; "archive anything from newsletters, flag anything from my direct reports, leave everything else" is a task your assistant can execute confidently without stopping to ask. Vague instructions are the most common reason a task ends up in needs_clarification — see Working with tasks for what that status means and how to respond to it.

Only install the skills you actually need

Every installed skill expands what your assistant can do — and what it might do unprompted. Org and team admins should periodically review installed skills against what people actually use, especially for anything with write access to external systems (calendars, email, CRMs).

Treat scheduling conflicts as information, not noise

When your assistant proposes a time and flags a conflict, that's it surfacing a real tradeoff, not being unhelpful. Read the reasoning it gives rather than always picking the first suggested slot — see How scheduling works for how it weighs competing constraints.

Keep connections current

A stale or revoked connection (an expired OAuth token, a calendar you no longer use) causes tasks to fail in ways that look like your assistant being wrong when the actual cause is upstream. Check Connections periodically, especially after changing passwords or revoking app access from a provider's side.