pm-feedback

name: pm-feedback description: Классифицирует пользовательский фидбек (Excel/CSV/текст) по 6 категориям, делает sentiment-анализ, кластеризацию тем, анализ трендов, триангуляцию по источникам, расчёт NPS и извлечение персон. На выходе — Top-10 болей с рекомендациями к действию. User-invoked only — do NOT auto-trigger. Triggers on /pm-feedback, "анализ обратной связи", "разбор отзывов", "анализ NPS", "analyze user feedback", "VOC analysis", "NPS analysis", "review analysis".

pm-feedback — User feedback analysis

Part of the Personal Corp framework — running a one-person business through AI agents. Structure raw feedback into a decision-driving insight report. Built-in classification, sentiment, theme clustering, NPS, trend analysis, source triangulation, and persona extraction.

Inputs

Mode: ≤ 20 items → close-read mode (item-by-item with detailed reading); > 20 → statistical mode (auto-classify + aggregated report).

Step 1 — Pre-process data

• Drop exact duplicates
• Merge near-duplicates (similarity > 90%), record merge count
• Ultra-short items (< 5 chars, no substance like "good"/"bad") → counted separately, not in deep analysis
• If a rating column exists (1-10 or 1-5 stars) → extract for NPS
• Identify source channel (in-app feedback, app store, support ticket, social media, etc.)

Step 2 — Classification

Six-category taxonomy:

When ambiguous (one item spans multiple), tag primary + secondary.

Step 3 — Sentiment analysis

Negative-intensity grading:

• Mild: calm dissatisfaction ("not very convenient")
• Medium: explicit disappointment ("very disappointed", "bad experience")
• Severe: threats ("I'll uninstall if not fixed", "I'll file a complaint") → high-priority handling

Step 4 — Theme clustering

Apply two methods to extract core themes.

Method A — Affinity mapping:

1. Split observations: decompose each feedback item into independent observation cards
2. Natural cluster: group by similarity without preset labels — let themes emerge
3. Name themes: label each cluster ("payment flow friction", "search results irrelevant")
4. Identify hierarchy: group small clusters under larger themes (e.g. "payment friction" + "long refund cycle" → "transaction experience")
5. Flag outliers: items that fit no cluster — possible early signals

Method B — Thematic coding:

1. Open coding: tag each item with descriptive labels ("slow load", "crash", "hidden entry point")
2. Axial coding: group descriptive labels into abstract themes ("slow load" + "crash" → "performance issues")
3. Selective coding: identify core themes and their relationships
4. Quantify frequency: count mentions and share per theme

Cluster output:

Step 5 — NPS analysis (if rating data exists)

• NPS = % Promoters (9-10) − % Detractors (0-6)
• Industry benchmarks: SaaS avg 30-40, consumer apps avg 20-30
• 5-star → 10-pt mapping: 5★=10, 4★=8, 3★=6, 2★=4, 1★=2

Step 6 — Trend analysis (if time data exists)

MoM (or WoW) change calculation:

• Aggregate by week or month per category
• Growth rate = (current − previous) / previous × 100%
• Watch for > 30% changes — flag as "needs attention"

Inflection-point detection:

• 3+ consecutive periods in one direction → established trend
• Sudden direction reversal → trigger investigation
• Correlate with external events: releases, campaigns, competitor moves

Trend output:

• Time-series description per category
• Mark significant changes + likely cause
• Early-warning: which metrics are deteriorating, which improving

Step 7 — Triangulation

When data spans multiple channels, cross-validate to lift confidence.

Method triangulation: same problem confirmed by different methods

• e.g. theme cluster says "slow load = top pain" → check if NPS detractors' open-ended answers also concentrate on performance

Source triangulation: same finding across channels

• App-store complaints + support tickets + community chatter all cite "crash" → high confidence
• Single-channel finding → tag "single-source, needs validation"

Time triangulation: persistence of the same problem

• 3 weeks consistent → systemic

• One-off → likely transient or already fixed

Confidence tiers:

Step 8 — Persona extraction

Identify typical user types from the feedback corpus.

Method:

1. Behavior cluster: infer user types (newbie / veteran / power user / occasional)
2. Need cluster: which users care about efficiency, which about experience, which about price
3. Sentiment cluster: loyal advocates / silent users / vocal complainers / churn-edge

Persona template:

[Persona name]: {one-sentence description}
- Typical traits: {usage frequency, focus, behavior pattern}
- Core need: {primary concern}
- Main pain: {recurring problem}
- Feedback style: {how they express}
- Estimated share: {% of feedback corpus}
- Quote: "{verbatim}"

Cap at 3-5 personas — more loses actionability.

Step 9 — Pain-point ranking

Pain priority = Frequency × Severity × User weight × Confidence

Sort descending; output Top 10.

Step 10 — Generate report

# User Feedback Analysis Report

**Period:** {date range}
**Total feedback:** {N} (after dedup: {M})
**Sources:** {channel list}

## 1. Classification
| Category | Count | Share | MoM change (if available) |
|---|---|---|---|

## 2. Sentiment
**Positive:** {X}% | **Neutral:** {Y}% | **Negative:** {Z}%
(Negative breakdown: mild {a} / medium {b} / severe {c})

## 3. Themes
| Theme | Sub-theme | Mentions | Share | Confidence |
|---|---|---|---|---|

## 4. NPS (if rating data)
**Score:** {n} (Promoters {X}% − Detractors {Y}%)
**Benchmark:** {above/below} industry by {Δ}

## 5. Trends (if time data)
- Significant rises: {category}, +{X}% MoM
- Significant drops: {category}, −{X}% MoM
- Inflection events: {description}

## 6. Top 10 Pain Points
| Rank | Pain | Freq | Severity | Confidence | Score | Quote | Recommendation |
|---|---|---|---|---|---|---|---|

## 7. Personas
<!-- 3-5 personas -->

## 8. Key Insights
<!-- Each insight: finding + data + confidence + meaning -->
1. {insight 1}
2. {insight 2}
3. {insight 3}

## 9. Improvement Recommendations
| Priority | Recommendation | Linked pain | Expected impact | Validation method |
|---|---|---|---|---|

## 10. Statistical Notes
- Classification confidence: {high/medium} (sample {N})
- Ambiguous classifications: {count}
- Triangulation coverage: {X%} of findings multi-source verified
- Validity: {sufficient sample / limited sample, results reference-only}

Quality bar

1. Classifications grounded; ambiguous items tag confidence
2. Insights backed by numbers; every insight cites a count
3. Recommendations actionable to feature level
4. Sample < 50 → tag "limited sample, results reference-only"
5. Stats computed via code for accuracy
6. Sentiment runs through calibration rules
7. Theme clusters MECE (mutually exclusive, collectively exhaustive)
8. Triangulation tier explicit per finding

Red lines

1. No over-extrapolation — 3 of 20 items mention X ≠ "many users say X"
2. Preserve verbatim — every pain point includes a representative quote for traceability
3. No fabricated trends — no MoM analysis without history
4. No invented personas — personas grounded in cluster results, not imagined

When input is incomplete

• < 10 items → close-read each; skip statistics (sample too small)
• No source/time info → analyze, but tag "missing source/time, recommend supplementing"; skip trend + triangulation
• Mixed languages → group by language, analyze separately
• Single-source → analyze, but tag "single source, recommend cross-channel validation"

Related skills

• /pm-prioritize — feature requests from feedback → RICE-rank
• /pm-prd — high-frequency requests → PRDs
• /pm-competitive — competitor mentions in feedback → enrich competitor study
• /pm-metrics — cross-validate feedback trends with product metrics