Human-centered collaboration.
In an AI-intensive session, one person becomes the AI operator while everyone else watches. These are the moves that stop that happening.
Rotating roles
Within each team, assign roles that rotate as the group moves through the phases. It takes thirty seconds to set up and changes who talks.
- Prompter
- Types or speaks to the AI. The obvious job, and the one that takes over the whole session if no one rotates it.
- Challenger
- Questions assumptions. AI tools fill gaps when they lack context, and someone has to watch for it.
- Synthesizer
- Captures what the humans said. This is the job that puts the "human" in human-centered.
- Keeper
- Holds the big picture. Teams follow a tangent and lose twenty minutes.
The human-first protocol
For each phase, require human work before any AI use. The format does this by default, but say it out loud anyway.
- Designer
- Five minutes of discussion and sharing observations before anyone opens a tool.
- Builder
- Sketch rough ideas on paper before asking AI to expand them.
- Prototype
- Make low-fidelity elements first, then use AI to improve them.
- Advocate
- Outline the key points yourselves before AI helps structure them.
Reflection checkpoints
Run a micro-reflection between prompts. Each takes under a minute and keeps the team in charge of the tool.
- What did the AI add, change or skip that surprised you?
- What did you know that the AI missed?
- How would you write that prompt differently next time?
Warning signs of over-reliance
Each of these looks like progress. If you see one, close the laptop and ask the team to explain their work to you.
- Designer
- AI generates the problem statement and the team can't explain why it matters.
- Builder
- AI produces a perfect journey map and the team has built no empathy.
- Prototype
- AI builds a complete prototype and the team can't iterate on it.
- Advocate
- AI writes the report and the team reads it out without conviction.
If a team cannot tell you, in their own words and without reading a screen, what problem they are solving and why it matters, they have outsourced the wrong part. Ten minutes with the laptops shut usually fixes it.