AI team
AI team
Asking one AI to do everything produces a narrow view. An AI team assigns each role its own persona and way of working. A commander breaks the request down, the specialists work, and the results are pulled back together.
The 17 role presets
Roles fall into four categories. Each one carries a persona and a playbook β a defined way of proceeding β plus a default mode that matches the work.
Core (useful on any team)
| Role | Responsible for | Default mode |
|---|---|---|
| π Commander | Breaking down the request, directing specialists, integrating results | Chat |
| 𧩠Planning & strategy | Goals into KPIs, and the order of moves | Chat |
| π Research | Turning a question into what needs finding out, with sources | Chat |
| βοΈ Writer | Defining reader and purpose, then writing | Article |
| π Reviewer | Checking by dimension: problem β reason β suggested fix | Chat |
Creative
| Role | Responsible for | Default mode |
|---|---|---|
| π£ Marketing | Positioning, funnel and calls to action | Chat |
| π¨ Designer | Tone, colour and layout | Image |
| π» Developer | Confirming requirements and constraints, then implementing | Code |
| π¬ Video director | Structure, shots, narration and music | Video |
| βοΈ Editor | Cutting the slack, pacing, order of information | Video |
Operations
| Role | Responsible for | Default mode |
|---|---|---|
| βοΈ Operations | Standardising and automating workflows | Chat |
| π§ Customer support | Empathy, resolution, turning it into an FAQ | Chat |
| π€ Sales | Hypotheses about customer problems, and value propositions | Chat |
| π Data analysis | Turning questions into metrics, decomposing causes, drawing implications | Chat |
Back office
| Role | Responsible for | Default mode |
|---|---|---|
| π Finance | Cost, revenue and margin structure; plan versus actual | Chat |
| π€ HR | Defining who you're hiring, plus evaluation and development | Chat |
| βοΈ Legal & compliance | Applicable law, disclosure wording, risk | Chat |
Ready-made line-ups
Rather than picking roles from scratch, start from a set that matches the job.
| Template | Roles included |
|---|---|
| Web production | Commander, research, marketing, designer, developer, reviewer |
| Video production | Commander, planning, writer, video director, editor, reviewer |
| Process improvement | Commander, research, operations, data analysis, reviewer |
| Customer response | Commander, support, writer, reviewer |
| Back office | Commander, finance, operations, legal, reviewer |
| Choose your own | Commander only β you add the rest |
Forming a team
Building it from the conversation you're already having carries the most context.
Open the thread where you've been discussing the work.
Choose π Create an AI team from this thread in the thread menu.
Team name, objective and suggested roles arrive pre-filled from the conversation. Adjust what needs adjusting.
On creation you land in the commander thread, with a draft of the first instruction in the composer.
Read it, then send β that's what starts the work.
You can also start empty from οΌNew β AI team.
Creating a team doesn't execute anything. It waits until you send the first instruction yourself.
Running and watching it
Send a request to the commander and it's split into stages and assigned. Each specialist works in a child thread.
- Progress is visible on the team progress view
- To halt everything, press Esc
- To consult the commander without mobilising the team, switch sending to Consult the commander
To hand an in-flight conversation over mid-way, use Hand over to the AI team in the send options.
Wrapping up
- π Finish: return results and disband β collects the team's output and closes it down
- β Return results to the parent β posts one child thread's conclusion upward
Custom agents
When no preset fits, build the specialist yourself. Manage them under Agents in the sidebar.
Create a new agent.
Give it a name, a one-line description and its intended audience.
Write its persona and expertise (who it is) and its method and rules (how it answers). This is what determines quality.
It becomes selectable from the role picker in any thread.
Mirror the structure of the built-in roles. Define who it is, number the steps it should follow, and specify the shape of the output β for example βproblem β reason β suggested fixβ.
Where to go next
- Your digital twin β teach it how you decide
- Growing context β how child threads and returning results work
- Skills β give roles a procedure to follow