Roll the week's frequent questions into a knowledge base
Operators explain the same thing twenty times a week. The agent turns the repeats into articles — the customer's wording, the answer from real threads.
Medium · 30 min · once a week
What you get
Draft base · 6 articles from 412 requests this week "Why is my order stuck in packing?" 41 requests · ready "Can I change the size after paying?" 28 requests · ready "When do I get my money back?" 23 requests → needs a person: operators quoted three different times
A sample on made-up data — your numbers will be your own.
Who it fits
- Operators answer the same question twenty times a week and write it out from scratch every time.
- A knowledge base exists, but it is written from inside the company and customers do not recognize their question.
When it won't work
The flow is small or every case is one of a kind — the articles end up describing single incidents.
How the agent does it
Set period and threshold
A week gives a live snapshot. Below five repeats random things slip in, so name your own threshold and the topics a lawyer has to clear.
Run the build
The agent groups requests by meaning and takes the answer from the operators' own replies. Gaps are flagged as discrepancies, never filled in.
Proofread before publishing
A knowledge base is read by customers with no operator nearby. A mistake spreads quietly — especially in prices and delivery times.
What you set
Connect Telegram, Avito and email. By hand: the period to look at, how many repeats make a question frequent, and topics that must not be published.
What you'll need
Starter prompt
Copy the prompt or open it straight in a chat with the agent.
Collect the recurring questions from tickets in Telegram, Avito and email for [period] and turn them into knowledge base articles. Group by the meaning of the question, not by matching words: "when will it arrive" and "where is my parcel" are one article. Take only topics that occurred at least [N] times, put the rest in a "rare" list. The title uses the customer's wording, the answer comes from real operator replies with a link to the source ticket. Where operators answered differently, do not pick the correct version yourself: show both answers and mark the article "needs a decision". Output articles with the fields question, answer, source, ticket count — in descending order of frequency.
Open in chatCheck the draft base. List: — articles built on fewer than [N] tickets; — groups where questions of different substance sit under one title; — answers where you filled in the missing part yourself instead of taking it from the correspondence; — topics where operators answered inconsistently; — answers with prices, deadlines and promotion terms — they need a review date; — what can be published and what should first be shown to a lawyer or the support lead.
A second prompt — the agent uses it to review its own work and show what's left for you.