Use when you need to know where value concentrates and which of those segments you can actually act on.
segment-value-profiling.md
You are profiling segments to find where value concentrates. You are not writing personas and you are not describing anyone in personality terms.
Segment data with counts and value: {{SEGMENT_DATA}}
How value is defined here: {{VALUE_DEFINITION}}
Dimensions available to cut by: {{AVAILABLE_DIMENSIONS}}
What we can do differently per segment: {{ACTIONABLE_LEVERS}}
Output:
1. A table: Segment | Share of users | Share of value under {{VALUE_DEFINITION}} | Value per user | Index against the average | Sample size, flagged where it falls under {{MIN_SEGMENT_SIZE}}.
2. The segments where share of value and share of users diverge most, with both numbers quoted from {{SEGMENT_DATA}}.
3. For the top three, the dimensions from {{AVAILABLE_DIMENSIONS}} that define them and whether each can be reached by a lever in {{ACTIONABLE_LEVERS}}. A segment we cannot reach is an observation, not a target. Say which are which.
4. Overlaps: segments that are largely the same people described two ways, named as pairs.
5. What is confounded here, for example a difference that is more likely tenure, acquisition channel, or product mix than the segment itself.
Rules:
- Do not report an index for any segment under {{MIN_SEGMENT_SIZE}} users. List those separately as too small to read.
- Averages hide skew. Where the value in a segment is carried by a handful of accounts, say so and give the figures.
- No causal language anywhere. These are differences, not effects.
Replace each placeholder with your own detail. The more specific you are, the less the model invents.
When do I use this rather than building personas?
When you need to know where value concentrates and which of those concentrations you can act on. Personas describe people. This quantifies share of value against share of users, then asks whether a lever exists to reach each segment. It keeps personality language out entirely, which is deliberate rather than dry.
What do I need in front of me?
Segment data with counts and value, an explicit value definition, the dimensions available to cut by, the levers you can pull per segment, and a minimum segment size. Cut by one dimension at a time first. Combined segments get small faster than people expect, and the index column then reports noise.
What comes back?
A concentration table with share of users, share of value, value per user and an index against the average, the biggest divergences quoted from your data, whether the top three are reachable by an actual lever, overlapping segments named as pairs, and what is confounded. The reachability check is the useful bit.
What is the mistake that costs me here?
Reading a difference as an effect. The prompt keeps causal language out and names likely confounders such as tenure, acquisition channel and product mix, because a high value segment usually describes who bought rather than why. Re-run it on a later period before committing budget, since concentration moves with the new customer mix.