Feedback Triage

Purpose

PMs drown in feedback. 200 Intercom tickets, 50 sales call notes, 80 NPS comments, a Slack channel of complaints. Manual triage takes a day and you still miss patterns.

This skill turns a raw dump into a ranked, themed report with enough structure to drive prioritization decisions and enough quotes to defend them.

The output is opinionated: it ranks themes by a composite score, not just frequency. A theme mentioned once by a strategic enterprise customer can outweigh 30 generic requests.

When to use

• Quarterly feedback review
• Post-launch: triage the inbound noise
• Onboarding: get a fast read on what customers actually want
• Pre-roadmap-planning: ground the strategy in real evidence
• Customer-success or sales-team feedback synthesis

When NOT to use

• For 5-10 feedback items - just read them
• For a single critical incident - dedicated triage, not statistical
• For real-time monitoring (use alerting tools)

Inputs

Required:

• Feedback items: paste, file (CSV / JSON / markdown), or path

Strongly improves quality:

• Customer metadata (segment, plan tier, ARR, account name) per item if available
• Business priorities or strategic themes the PM cares about
• Time window the feedback covers

Default assumption: every item is one unit unless metadata says otherwise. With metadata, items can be weighted.

Process

Phase 1: Normalize

• Deduplicate near-identical items
• Standardize format (each item: id, text, source, date, customer metadata if any)
• Flag items that are not feedback (sales pitches, internal notes, irrelevant chatter) and exclude from clustering
• Report normalization stats: original count, deduped count, excluded count

Phase 2: Theme extraction

Cluster items by underlying job-to-be-done, not surface words. "I need an export to CSV" and "give me a way to get this data into Excel" are the same theme.

For each theme:

• Theme name (concrete, not generic - "Bulk edit roles in user admin" not "Better UX")
• Underlying JTBD ("When I'm onboarding a new team, I want to assign roles in bulk, so I can finish setup faster")
• Item count
• Sample quotes (3, diverse)

Aim for 8-15 themes. More than 20 means theming is too granular.

Phase 3: Score each theme

Four sub-scores (1-5 each):

• Volume: how many items, normalized
• Severity: how blocking is this for affected users
• Strategic fit: does this align with stated priorities
• Customer weight: are these high-value accounts (uses metadata if provided)

Composite priority = average of the four (1-5).

Report each theme with all four sub-scores plus composite. Do not hide the inputs - reviewers will want to challenge specific dimensions.

Phase 4: Sentiment slice

For each theme:

• % positive / neutral / negative
• Tone shift over time if time-window data available
• Flag themes where sentiment is sharply negative (escalation candidates)

Phase 5: Surprises and orphans

• Surprises: themes the PM probably didn't expect (compare against stated priorities)
• Orphans: items that didn't cluster - sometimes these are the most interesting (early signals)
• Conflicts: where customer segments want opposite things

Output

# Feedback Triage Report

## Stats: original / deduped / excluded
## Top 5 themes (executive summary)

## Full theme table
| # | Theme | JTBD | Items | Volume | Severity | Fit | Weight | Composite | Sentiment |

## Detailed theme breakdowns
[For each theme: name, JTBD, sample quotes, scoring rationale]

## Surprises
## Orphans worth attention
## Conflicts across segments
## Recommended next actions (3)

The output should be paste-into-Notion ready. Tables, not paragraphs, for the structured data.

Common failure modes to avoid

• Theme inflation: 30 themes is not analysis, it's transcription
• Generic theme names: "Better mobile experience" is not actionable. Be specific.
• Hiding scoring inputs: if you only show composite, reviewers can't challenge it
• Ignoring metadata: feedback from 5 enterprise customers worth $2M ARR is not the same weight as 50 free-tier users
• Pretending the data is clean: report the noise honestly. PMs need to know how much they're trusting.

Tools strategy

If the input is a structured file (CSV, JSON), parse it programmatically. If pasted text, normalize first, then process. For 500+ items, batch the clustering and consolidate.