Feedback Triage

Feedback Triage

Turn a feedback dump into a ranked, countable picture of what's actually hurting users — with the receipts attached.

Before triaging

1. Confirm the source and time window. Reviews skew angry; sales notes skew deal-driven; NPS skews extremes. Name the skew in the output.
2. Ask if there's a segmentation field worth cutting by (plan tier, platform, tenure). A complaint from 40% of enterprise accounts ranks differently than 40% of free users.
3. For large files (>200 items), use code (Python/pandas) to count rather than eyeballing — exact counts, not impressions.

Workflow

1. Read/parse everything. For structured exports, load programmatically; dedupe obvious repeats from the same user.
2. Tag each item with: theme, severity (blocker / major friction / minor / request), and segment if available. Build the taxonomy bottom-up from the data — don't force items into preconceived buckets. Cap at ~10 themes; merge or use "Other" beyond that.
3. Count honestly. Report both raw counts and % of total. Never say "many users" when the number is 7.
4. Rank by pain, not just volume. Rank = f(count, severity, segment value, trend). A low-volume blocker for paying customers can outrank a high-volume cosmetic gripe — show the reasoning when rank diverges from raw count.
5. Pull 2–3 verbatim quotes per top theme. Quotes travel further in orgs than counts do.
6. Recommend a next action per top theme: fix now / investigate / add to backlog with link / communicate workaround / no action (with reason).

Output format

# Feedback triage — [source], [window]
**Volume:** [N] items ([deduped from M]) · **Known skew:** [source bias]

## Headline
[One sentence: the single most important thing this data says.]

## Themes by pain rank
| # | Theme | Count | % | Severity | Segment concentration | Trend |
|---|---|---|---|---|---|---|

## Top themes in detail
### 1. [Theme]
[Counts, quotes with source IDs, why it ranks here, recommended action + owner suggestion]

## Divergences
[Where pain-rank ≠ volume-rank, and why]

## Watch list
[Low-volume but new or fast-growing signals]

## Method notes
[Taxonomy decisions, dedupe rules, anything ambiguous in the tagging]

Quality bar — self-check

• Every count is exact (computed, not estimated) for structured inputs.
• Quotes are verbatim with a traceable ID.
• Source skew is named in the header.
• Rank divergences from volume are justified, never silent.
• "No action" is used where honest — recommending action on everything is triage failure.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: ramanbamba
• Source: ramanbamba/10x-pm
• License: MIT

Install and usage instructions live in the source repository linked above.