Skip to main content
AI coworkers work best when you treat them like a capable new hire: give them a clear job, the context they need and feedback when they get something wrong.

​
Give each coworker a clear job

One coworker per team or job works better than one coworker that does everything. “Recruiting coordinator” or “Customer success analyst” gives the coworker a focus, and makes it easy for your team to know who to ask. Put the job description in its AGENTS.md: what it’s responsible for, who it works with, and what it should never do without asking.

​
Ask for outcomes, with context

Say what you want done, where the information lives, and what the result should look like.
Look at our pipeline.
Useful details to include:
  • Sources. Which tool, channel, database or file to use.
  • Scope. Time range, team, region or record type.
  • Format. Table, bullet list, short summary, draft email or artifact.
  • Destination. Reply in the thread, post in a channel, DM someone or send an email.
  • Approval. Whether it should act or show you a draft first.

​
Draft first, act second

For anything that leaves Vybe, such as emails to customers, CRM updates or messages in public channels, ask for a draft first:
Draft a follow-up email to everyone who attended yesterday’s webinar. Show me the draft before sending anything.
Once you trust the output, let the coworker act directly. You can also limit what it’s allowed to do: set a connection to Read only in the coworker’s Connections section until you’re comfortable. See Capabilities.

​
Turn repeated requests into automations

If you ask for the same thing every week, make it a scheduled task:
Every Friday at 4pm, summarize this week’s closed tickets in Intercom by category and post it in #support-leads.
If the work should start when something happens, use a trigger instead.

​
Correct it, then ask it to remember

When a coworker gets something wrong, tell it what you expected and ask it to remember:
That’s not quite right: “active customers” means accounts with a paid invoice in the last 90 days, not anyone who logged in. Remember that for next time.
The coworker saves a note, and it will use that definition in future conversations. For multi-step procedures, ask it to save a skill:
Save the way we just prepared the board update as a skill, so you can do it the same way next month.
Review what it has learned in Knowledge, and lock notes it shouldn’t change.

​
Keep conversations focused

Start a new thread for each new topic. Long threads that mix several jobs are harder for the coworker to follow and use more credits. In the web app, remember that everyone who can access the coworker can read its chats.

​
Check its work

Open Usage & Activity → Activity on the coworker’s page to see each run: what it was asked, how it answered and every step it took. This is the quickest way to understand a surprising result. See Usage and activity.

​
Pick the right model

Auto works well for most coworkers. If a coworker handles hard analysis or long multi-step jobs, try a stronger model or higher effort. For simple, high-volume work, a faster model saves credits. See AI models.