SOCIAL

Turn a pile of mentions into a digest

Use when a listening export runs to hundreds of rows and you need themes rather than a wall of text.

social-listening-digest.md

Download .md

You are reading a social listening export and reporting what is in it. You are not scoring sentiment and you are not drafting replies.

Mentions export, one per line with source and date: {{MENTIONS_EXPORT}}
Period covered: {{PERIOD}}
What we currently believe people say about us: {{CURRENT_ASSUMPTIONS}}
Products or topics we track: {{TRACKED_TOPICS}}

Output four numbered sections.

1. Themes as a table: Theme | Mentions in it | Representative quote, verbatim | Which of {{TRACKED_TOPICS}} it touches | New or known. A theme needs at least three separate mentions in {{MENTIONS_EXPORT}}; anything smaller belongs in section 3.
2. Assumption check: each item in {{CURRENT_ASSUMPTIONS}} marked supported, contradicted or absent, with the mentions that decide it.
3. Weak signals: single mentions worth watching, each with what would confirm it.
4. Needs a human: complaints, factual errors about us, and mentions from accounts with reach, ordered by urgency, with who should handle each.

Quote exactly and never merge two people's words into one quote. Do not report a percentage or a change against an earlier period unless {{PERIOD}} contains the comparison data. Where volume is driven by one thread or one account, say that instead of reporting it as a theme.

Fill in before running

Replace each placeholder with your own detail. The more specific you are, the less the model invents.

{{MENTIONS_EXPORT}}
{{PERIOD}}
{{CURRENT_ASSUMPTIONS}}
{{TRACKED_TOPICS}}

Getting a better result

Write your assumptions down before running it; checking them afterwards is what keeps the exercise honest.
Strip your own brand account and staff mentions out first, or the largest theme will be you talking.
Keep the weak signals between runs - a single mention that reappears three months running is a real theme.

Questions about this prompt

When should I use this rather than the sentiment score from the tool?

When the export runs to hundreds of rows and reading it end to end is a day nobody has. A sentiment score collapses everything into a number that hides what people actually said. This prompt is told not to score sentiment and to quote verbatim instead, so you can read the wording yourself.

What do I need in front of me?

The export with source and date per line, the period, your tracked topics, and your current assumptions written down before you run it. That last part is what keeps the exercise honest. Strip your own brand account and staff mentions first, or the largest theme will be your team talking.

What comes back, and which section changes anything?

A themes table needing three separate mentions each, an assumption check marking every belief supported, contradicted or absent, weak signals, and mentions needing a human. The assumption check is the section that changes what you do, because a contradicted belief is the only finding here that forces a decision.

What is the mistake?

Reporting one thread as a theme. The prompt is told to flag volume driven by a single account or conversation, and that line is easy to skip on the way to a slide. Carry the weak signals between runs too: a single mention that returns three months running has become a real theme.