All use cases
ProductAnalysis

Find the top 5 user pain points this quarter

"Users are complaining" gets you nowhere in a board meeting. The agent counts how many people each problem hit and how many of them never came back.

Medium · 30 min · once a quarter

What you get

Top pains, Q2 — 3,140 tickets
1. Slow delivery to the regions — 412 people
   did not return within 90 days: 118 (29%)
2. Card payment failure — 187 people
   did not return: 71 (38%)
3. Sizing confusion — 260 people
   effect not computed: 190 tickets could not be linked
   to a buyer.

A sample on made-up data — your numbers will be your own.

Who it fits

  • Quarterly planning where you must pick the costliest problem, not the most frequent one.
  • A conversation with management where lost-customer numbers land and adjectives do not.

When it won't work

Tickets cannot be linked to a person's later purchases — you get frequency and no money effect.

How the agent does it

1

Set quarter and rules

Say what counts as a lost customer and over what window, or the agent invents its own rule. Name the attribute that links a ticket to a person.

2

Run the analysis

The agent groups recurring problems, counts how many people each one reached and looks separately at what those people did next.

3

Split counted from guessed

Ask for two lists: pains where the effect is computed from data, and pains where it is only asserted — with the sample size under every number.

What you set

Telegram, Ozon, Wildberries. By hand: the quarter boundaries and your own definitions of churn and of a repeat purchase.

What you'll need

Telegram
Ozon
Wildberries

Starter prompt

Copy the prompt or open it straight in a chat with the agent.

Prompt for the agent

Go through the Telegram tickets and the Ozon and Wildberries reviews for [quarter] and identify the five main user pain points. Rank them not by mention frequency but by cost: how many users are affected and what happened to them afterwards — [left / did not buy again / reduced their basket]. State the effect on churn and repeat purchases only where you computed it from data you actually saw: attach which rows the figure was built from and how many people are in the sample. Where linking a ticket to subsequent behavior did not work out — write "effect not measured" and leave the pain point in the list by frequency, instead of grading it with words like "probably has an effect". Output five cards: the wording of the pain point, how many users are affected, [3] quotes, what happened to them afterwards — and, as a separate list, the hypotheses that the data does not confirm.

Open in chat
Check the result

Check the top pain points. List: — pain points where the churn effect is computed and pain points where it is merely asserted — as two lists; — the sample size behind every effect figure and how many people could not be linked to purchases; — pain points that made the top thanks to one noisy channel; — complaints you merged into a single pain point, and by which signal; — cases where the number of users grew along with the complaints and the shares did not change; — which conclusions are verifiable against the data and which are your hypothesis; do not mix hypotheses into the top.

A second prompt — the agent uses it to review its own work and show what's left for you.

Related use cases