CSAT / NPS Analysis Skill

A satisfaction score on its own is a vanity number — the value is in why it's that number and what to do. This skill computes the score correctly (NPS is %promoters − %detractors, not an average), reads the verbatims for the themes driving promoters and detractors, and turns it into a prioritised action list — so a survey becomes a roadmap, not a slide.

Required Inputs

Ask for these only if they aren't already provided:

• The metric & data — NPS (0–10 ratings), CSAT (e.g. 1–5 or % satisfied), or CES; the response counts/distribution.
• The verbatims — open-text comments (the gold; paste what you have).
• Context — segment, time period, and the prior score for trend.

Output Format

[CSAT / NPS / CES] Readout: [segment, period]

1. The score — computed (use the helper for NPS/CSAT): the headline number, the distribution (promoters/passives/detractors for NPS), the trend vs. last period, and the benchmark (industry/your target). State the formula — NPS is a net of percentages, not an average.

2. What's driving it — theme the verbatims:

• Promoters love: the 2–3 recurring reasons people rate high (protect/amplify these).
• Detractors hurt by: the 2–3 recurring pains (these are your fix list).
• Passives need: what would move them up. Quote a representative comment per theme.

3. Segments — where the score is notably worse/better (plan, tenure, channel), if the data allows — the average hides this.

4. Actions — prioritised: the highest-frequency × highest-impact detractor themes first, each with an owner and the metric it should move. A score with no actions is wasted.

Programmatic Helper

scripts/nps.py (stdlib only) computes NPS / CSAT from the rating distribution:

# NPS from 0-10 counts (11 numbers, ratings 0..10):
python3 scripts/nps.py nps 12 5 8 ... 
# CSAT % satisfied (ratings 4-5 on a 1-5 scale):
python3 scripts/nps.py csat 2 3 10 40 55
python3 scripts/nps.py nps "...counts..." --json

Quality Checks

• [ ] NPS is computed as %promoters − %detractors (not an average of scores)
• [ ] The distribution and trend vs. last period are shown, plus a benchmark/target
• [ ] Verbatims are themed into promoter/detractor drivers, with a representative quote each
• [ ] Segment differences are surfaced where the data allows (the average lies)
• [ ] Ends with prioritised, owned actions tied to the biggest detractor themes

Anti-Patterns

• [ ] Do not average NPS ratings — it's a net of percentages; averaging gives a meaningless number
• [ ] Do not report the score without the why — the verbatims are where the action is
• [ ] Do not ignore passives — they're the cheapest group to convert into promoters
• [ ] Do not stop at the score — an analysis with no prioritised action changes nothing
• [ ] Do not trust a tiny sample — flag low n; a 12-response NPS swing is noise, not a trend

Based On

Voice-of-customer practice — correct NPS/CSAT/CES computation, verbatim theming, and action prioritisation.