Skills CX operations
Designing support surveys that measure something
A satisfaction score is a statistic about the people who chose to answer. Most survey programmes treat it as a statistic about customers. The gap between those two things is usually larger than any real change in service quality — which is why CSAT moves for reasons nobody can explain, and why it so often fails to predict churn.
npx rulebase-skills install cx-survey-designWhen to use it
Reach for this when someone says any of these — they are the phrases the skill itself triggers on:
- “design a CSAT survey”
- “our CSAT doesn't match reality”
- “should we use NPS”
How it works
The method, in the order the skill runs it. The full procedure — tables, worked examples and the edge cases — is in the skill itself.
Step 1: pick the right instrument
NPS after a support contact is the most common mis-specification in CX. A customer's willingness to recommend your company is driven mostly by the product, price, and their overall history — not by whether one agent was helpful. Using it transactionally holds agents responsible for things they do not control, and adds so much noise that real signal disappears. If you must report NPS, keep it as a separate relationshi
Step 2: fix solicitation before anything else
Nobody may choose who gets surveyed. If agents can trigger, suppress, or influence survey sends, the programme is measuring agent selection and nothing else. This is not a minor bias — it is fatal, and it is common.
Step 3: choose a scale and then leave it alone
A scale change breaks the time series permanently. There is no valid conversion. If you must change, run both in parallel for a period and report the break explicitly in every chart thereafter. Treat this as a one-way door.
Step 4: decide when to ask, and know what that decides
Both are valid. They are not comparable, and averaging them produces a number that means nothing. Pick one per programme and document it.
Step 6: measure your response bias
This is the step almost everyone skips, and it is the one that determines whether the score means anything.
Step 7: compare fairly, or don't compare
Comparing CSAT across teams, channels, or periods is only valid when response rates are comparable. A team with a 30% response rate and one with 10% are different samples, and the difference in their scores may be entirely sampling.
Step 8: sample size, honestly
Same arithmetic as any proportion. For a top-2-box rate p from n responses, the 95% interval is roughly ±1.96·√(p(1−p)/n).
Troubleshooting
CSAT is high but customers are churning — check response bias first. This is the classic signature of a self-selected sample.
Related skills
Free and open source, and vendor-neutral — it reads the conversations from whichever helpdesk you already run. Browse all 149 skills · connect your helpdesk over MCP · source on GitHub
