How fast do prompting habits form?

Within roughly five sessions, per the study Nielsen reports. Phrasings tried in the first five sessions recurred 5 to 50 times more often later, and users who typed a greeting early were 49 times more likely to keep doing it months on. I have read his account, not the paper.

What is a paired shipping session?

Ninety minutes, two people, one real artifact. A driver who has shipped something similar with the tool shares a screen with a rider who has not, and they build something for the rider's team. It is not a tutorial. If the artifact is not in the team's workflow within 48 hours, it does not count.

Should the product correct a new user's prompts automatically?

Carefully, if at all. Nielsen suggests showing a user's prompt beside an improved one. My experience with an agent that intervened in first sessions is that correct help can still read as surveillance, and activation dropped. Persistent controls the user chooses to use are the safer way to show alternatives.

What about people who have used the tool for months?

Wait for a major model release. In the study, the 246 users who moved from GPT-3.5 to GPT-4 started exploring again like novices. That is the moment to rerun the paired sessions for everyone.

When Should You Train a Team on a New AI Tool?

When should you train a team on a new AI tool?

THE SHORT ANSWER

In each person's first week with the tool, and again when a major model ships. A study Jakob Nielsen reported on October 9, 2026 (Shengqi Zhu and co-authors, 139,535 ChatGPT sessions from 7,955 users) found that people settle on 2 to 4 personal prompt templates within roughly 5 sessions and then stop inventing, while the tasks they bring keep growing. The data is correlational and from 2023 and 2024 logs. A recurring training slot reaches most people long after that window. I audited AI office hours in seven product orgs and found high attendance and no measurable change in what shipped. What I would run is a 90-minute paired shipping session inside the first five sessions: one person who has shipped with the tool and one who has not, building a real artifact for the newcomer's team, counted only if it is in that team's workflow within 48 hours. In the product itself, keep alternatives visible and let the user reach for them. My onboarding agent that noticed where new users got stuck and offered help lowered activation against the control group, because people read it as being watched.

Most rollouts hand out licenses first and schedule enablement later. The five-session finding says later is the expensive part. By the time the training happens, each person has two or three ways of asking that work well enough, and people rarely switch methods once one works.

So tie the help to the person's start date, not to the calendar. The cheapest people to help are the ones who got access this month.

The study, my failed onboarding agent, and what to change in a product's first run are in Five Sessions, Then the Prompting Habit Sets. The session format is from Kill the AI Office Hours. They're 2026's Agile Transformation.

Last reviewed 2026-10-10 · 1 min read

When should you train a team on a new AI tool?

SOURCES

THE LONG VERSION

RELATED ANSWERS