feedback-analysis

Feedback Deep Dive

Analyze customer feedback using Pendo's feedback tools to discover themes, extract actionable insights, and identify at-risk customers. This skill orchestrates three Pendo feedback tools — Pendo:generate_feedback_topics, Pendo:get_feedback_insights, and Pendo:get_feedback_items — into a comprehensive analysis. It also uses Pendo:list_all_applications to resolve subscription IDs and Pendo:searchEntities to look up account IDs by name.

When to Use This Skill

Any time the user wants to understand customer feedback: broad themes, specific topic deep-dives, risk identification, account-level feedback, or feedback filtered by type/alert/timeframe.

Parameters

The user's request may include any combination of these filters. If they don't specify, use sensible defaults (last 90 days, no filters).

• search_term: A topic to search for (e.g., "onboarding", "performance", "mobile app"). Uses semantic similarity matching, not exact string matching, so conceptual matches will be found.
• account_id: Narrow to a specific account's feedback
• feedback_type: One of: Product Enhancement Request, Product Issues, Pain Point, Positive Product Feedback, Competitor Weakness, Competitor Strength
• alerts: Filter by alert flags: Churn Risk, High Frustration, Blocker to Sale
• account_type: Customer, Prospect, or Churned
• timeframe: Date range to analyze (default: last 90 days — startDate 90 days ago, endDate today)

Prerequisites

Before running feedback queries, you need the user's Pendo subscription ID. Use Pendo:list_all_applications to get it. If the user has multiple subscriptions, ask which one to use.

Workflow

Step 1: Resolve Filters

Parse the user's request into concrete filter values. Calculate startDate/endDate from any timeframe mention (e.g., "last quarter" → startDate: 2025-10-01, endDate: 2025-12-31). If the user mentions an account by name but you don't have the ID, use Pendo:searchEntities with itemType ["Account"] to find it.

Build a feedbackFilters object that will be reused across all three feedback tools:

{
  "startDate": "YYYY-MM-DD",
  "endDate": "YYYY-MM-DD",
  "similaritySearchTerms": ["topic if provided"],
  "accountIds": ["id if provided"],
  "feedbackTypes": ["type if provided"],
  "alerts": ["alert if provided"],
  "accountTypes": ["type if provided"]
}

Key distinction on search terms:

• Use similaritySearchTerms when the user describes a topic conceptually (e.g., "feedback about performance" → ["performance"]). This does semantic matching and will find related feedback even if it doesn't contain the exact word.
• Use exactMatchSearchTerms only when the user explicitly wants exact phrase matching (e.g., "feedback that mentions the word 'latency' exactly").

Step 2: Generate Feedback Topics

Call Pendo:generate_feedback_topics with the feedbackFilters. This clusters all matching feedback into AI-generated themes and returns topic names, descriptions, and counts. This gives you the high-level landscape of what customers are talking about.

This is a slow-running tool — let the user know you're working on it.

Step 3: Get Insights and Raw Feedback (Parallel)

Run these in parallel since they're independent:

Insights: Call Pendo:get_feedback_insights with the same feedbackFilters. Returns distilled, actionable insights — each with a summary, explanation of why it matters, and a supporting quote from actual feedback.

Raw Feedback: Call Pendo:get_feedback_items with the same feedbackFilters. Returns up to 30 actual feedback items with titles, descriptions, account/visitor info, types, and alerts.

Risk Signals: Call Pendo:get_feedback_items again, but add alert filters for ["Churn Risk", "High Frustration", "Blocker to Sale"] (merged with any existing filters). This surfaces the most urgent feedback that needs attention. Skip this call if the user already filtered to a specific alert type, since it would be redundant.

Step 4: Synthesize the Analysis

Combine all results into a coherent report. The structure below is a guide — adapt it based on what data came back and what the user asked for. If the user asked a narrow question (e.g., "any churn risk feedback?"), don't pad the response with irrelevant sections.

Output Format

Feedback Analysis Report

Scope: Summarize what filters were applied (e.g., "All customer feedback about onboarding, last 90 days") Period: The date range analyzed Total Feedback: Count from raw feedback results (note if capped at 30)

Top Themes

Present the topic clusters from generate_feedback_topics as a table:

#	Theme	Description	Count
1	{topic}	{description}	{count}

Key Insights

For each insight, present:

{insight summary}

"{supporting quote}"

{Why this matters / explanation}

Risk Alerts

Group the risk-flagged feedback by alert type. For each, show the account name, a brief summary of their feedback, and the date. This section is critical — churn risks and high frustration should be immediately visible. If there are no risk alerts, say so (that's good news worth reporting).

Feedback Breakdown by Type

Summarize counts by feedback type (Product Enhancement Request, Product Issues, Pain Points, Positive Feedback, Competitor mentions) based on the raw items retrieved.

Recommendations

Based on the themes, insights, and risk signals, provide 2-4 actionable recommendations. Connect each recommendation to specific evidence from the feedback. For example, if multiple accounts mention slow report generation, recommend investigating performance and name the accounts affected.

Tips for Great Analysis

• If searching for a topic returns no results, suggest the user broaden their search or try related terms. Semantic search is forgiving, but very niche terms might not match.
• When presenting quotes from feedback, keep them brief and impactful — they add credibility to insights.
• If a specific account appears across multiple risk categories, call that out explicitly — it's a strong signal.
• For account-specific analysis, consider mentioning the account type (Customer vs Prospect) since the appropriate response differs.
• The raw feedback endpoint caps at 30 items. If 30 items are returned, note that there may be more matching feedback beyond what's shown.