The lessons learned thread

The thread, with margin notes

1. 7 things we learned from 4 months of beta testing an AI research assistant with analysts. Some of these cost us weeks:

   A numbered promise. "Cost us weeks" signals the lessons were paid for, which makes them worth saving.

2. 1. Analysts do not want answers. They want sources. Our first version wrote beautiful summaries. Usage dropped until we put the citation before the sentence, not after.

   Each lesson is a claim followed by the story behind it. The claim alone would be a platitude.

3. 2. Speed beats depth for the first query. A rough answer in 5 seconds kept people. A perfect answer in 90 seconds lost them.

   Contrast in two short lines. Easy to quote, easy to argue with, both good for reach.

4. 3. "Export to slides" was requested by everyone and used by almost nobody. Ask what people did last week, not what they want.

   A counterintuitive lesson that applies to any product. This is the one other founders repost.

5. 4. The best onboarding was a blank page with one example query. Our 6 step tour had a lower completion rate than no tour at all.

   Specific enough to be useful tomorrow. Avoid lessons like "listen to users" that teach nothing.

6. 5. Trust breaks on one wrong number. One hallucinated figure in a client memo and a tester stopped using it for a week. Now every number links to its source cell.

   Admitting the classic AI failure head on builds more trust than pretending it never happens.

7. 6. Teams adopt faster than individuals, if one person champions it. 7. Analysts talk to each other constantly. Our best signups came from a single shared screenshot.

   Combining the last two keeps the thread tight. Lesson 7 quietly argues for posting on X at all.

8. We launch in January. If you are an analyst and want to try it before then, the waitlist is below. And if you have a lesson I missed, I want it.

   Launch date, soft ask, and an invitation to add lessons, which fills the replies.

Next thread