--- name: voice-of-customer-rollup source: https://app.decimal.ai/s/voice-of-customer-rollup@1/SKILL.md source_sha256: 3f366fc35d7b --- # Voice-of-Customer Rollup ## What this does Takes a batch of raw customer feedback from one or more channels — support tickets, NPS comments, app-store reviews, survey open-text, sales-call notes — and produces a **fixed digest**: a list of themes, each **ranked**, each carrying a **count** and **one quote copied exactly from the feedback**, with **no theme that the feedback did not actually raise**. Handed a pile of comments and asked "what are people saying?", the base model writes a smooth paragraph. It groups loosely, **paraphrases** the customer's words into its own, attaches no counts, orders themes by whatever it mentioned first, and will happily add a tidy conclusion ("overall, sentiment is positive") that no single comment supports. This skill replaces that with a digest an operator can act on: themes ordered by real weight, each backed by a count you can trust and a customer's **own words**, and nothing invented. ## When to use / when NOT to **Use when:** the input is two or more pieces of customer feedback and the ask is to consolidate them — "roll these up", "what are the top themes", "summarize this feedback", "what are customers telling us across these". **Do NOT use when:** scoring one account's churn risk (that is a rubric task); writing a quarterly business review; drafting a reply to one customer; or doing sentiment on a single comment. This skill is for the many-into-a-digest move. ## The rule Every rollup obeys all six. The first three are what the base model gets wrong most. 1. **Quote verbatim — never paraphrase.** Each theme is anchored by **one representative quote copied word-for-word from an actual feedback item**. Do not clean up grammar, soften wording, merge two comments into one quote, or write a quote in your own voice. If you must shorten, cut only from the ends and mark the cut with `…` — the words you keep stay exactly as the customer wrote them. A quote a reader cannot find verbatim in the input is a fabrication. 2. **Invent no theme.** Every theme must be one that **at least one feedback item actually raises**. Do not add a theme because it is common in the industry, because you expect it, or to round out the list. If the feedback never mentions pricing, "pricing concerns" is not a theme. When in doubt, drop it. 3. **Count honestly.** Each theme carries a **count** = how many distinct feedback items raise it. The count reflects only items in the input. One item may support two themes; one item never counts twice toward the same theme. Do not estimate, round to a "roughly", or inflate a single mention into "several". 4. **Rank by frequency × severity.** Order themes by their weight, where weight combines **how many items raise it** (frequency) with **how much each instance hurts** (severity — a churn threat or blocked workflow outranks a mild annoyance). A theme raised 3 times where users say they will cancel outranks a cosmetic nit raised 5 times. State each theme's severity so the ranking is auditable; when two themes tie on weight, the higher-frequency one goes first. 5. **One item, its real theme.** Assign each feedback item to the theme(s) it genuinely raises. Do not force an off-topic comment into the nearest bucket to lift a count, and do not split one coherent complaint across three themes to pad the list. 6. **No editorial conclusion.** Do not append an overall-sentiment verdict, a recommendation, or a "customers love us but…" wrap-up unless the ask explicitly requests one. The digest is the themes and their evidence; a summary line the feedback does not collectively support is exactly the invention Rule 2 forbids. ## Output shape For each theme, in ranked order: ``` 1. — mentions · severity: "" ``` Lead with the highest-weight theme. Keep the label short and descriptive of what customers said, not of how you feel about it ("Login fails after password reset", not "Frustrating auth problems"). ## Worked example **Feedback given** (6 items): - [1] "Reset my password and now I can't log in at all — completely locked out." - [2] "The new dashboard is gorgeous, love the dark mode." - [3] "After a password reset the login just spins forever. Had to email support." - [4] "Export to CSV has been broken for a week, this is blocking our month-end close." - [5] "Wish the date picker remembered my last range, minor thing." - [6] "Password reset locked me out too, support took two days to fix it." **BEFORE (base default — paraphrased, uncounted, editorialized):** > Customers are generally happy with the new design but some are running into login trouble > after resetting their passwords, and a few mentioned issues with exporting and small UI > preferences. Overall sentiment is mixed but leaning positive. No counts, quotes reworded, themes ordered by nothing, and a sentiment verdict nobody said. **AFTER (this skill — ranked, counted, verbatim, nothing invented):** > 1. Password reset locks users out — 3 mentions · severity: high > "Reset my password and now I can't log in at all — completely locked out." > 2. CSV export broken — 1 mention · severity: high > "Export to CSV has been broken for a week, this is blocking our month-end close." > 3. New dashboard / dark mode praised — 1 mention · severity: low > "The new dashboard is gorgeous, love the dark mode." > 4. Date picker should remember last range — 1 mention · severity: low > "Wish the date picker remembered my last range, minor thing." The reset lockout leads on frequency (3) and severity (high); the export bug is a single mention but ranks above the cosmetic items because it blocks a workflow. Every quote is one customer's exact words. No pricing theme, no sentiment verdict — neither is in the feedback. ## Edge cases - **A theme is raised once but is severe** (a blocker or churn threat). Keep it and let severity carry it up the ranking — do not drop it just because frequency is 1. - **Mixed sentiment on the same feature** (some praise it, some hate it). Split into two themes with their own counts and quotes, or name it one theme and note the split — do not average the sentiment away. - **One comment raises several distinct issues.** Assign it to each relevant theme; it counts once toward each, never twice toward one. - **A vague comment** ("meh, not great") that names no issue. It supports no specific theme; do not manufacture one from it. - **The ask requests a recommendation or headline.** Then a summary line is in scope — but it must still follow only from the themes present, with no figure or claim the feedback lacks. ## Do / Don't - **Do** copy the customer's exact words as the quote. - **Don't** paraphrase, tidy grammar, or write the quote in your own voice. - **Do** rank by frequency × severity and state each severity. - **Don't** order themes by mention-order or by count alone. - **Do** keep a rare-but-severe theme. - **Don't** add a theme the feedback never raised, or a sentiment verdict it never supports. ## Common mistakes (the base's defaults) - Paraphrasing quotes into the model's own words instead of copying them verbatim. - Omitting counts, or writing "several / a few" instead of a number. - Ordering themes by whichever was mentioned first, ignoring severity. - Padding the list with an expected-but-absent theme (pricing, onboarding) nobody raised. - Tacking on an "overall sentiment is positive" verdict the individual comments don't support. ## Quick checklist - [ ] Every theme has a count = distinct items raising it. - [ ] Every theme is anchored by a quote copied word-for-word from an input item. - [ ] Themes ranked by frequency × severity, with each severity stated. - [ ] No theme appears that no feedback item raised. - [ ] No overall-sentiment verdict unless the ask requested one.