customer-feedback-ingestion

name	customer-feedback-ingestion
description	Ingests customer feedback from configured channels (GitHub issues, support tickets, app reviews, Slack/Discord). Categorizes, deduplicates, and produces a structured feedback digest. Does not interact with customers.
agents	{"primary":"business-analyst","supporting":["research-agent"]}
inputs	["discovery-config.yaml (channel configuration)",".factory/discovery/feedback-state.yaml (last ingestion timestamps)"]
outputs	[".factory/discovery/feedback-digest-YYYY-MM-DD.md",".factory/discovery/feedback-state.yaml (updated)"]

Customer Feedback Ingestion

Ingests customer feedback from configured channels, categorizes it, deduplicates similar items, and produces a structured feedback digest. This is a read-only process -- the factory does not interact with customers directly.

Trigger

• Scheduled -- daily or as configured in discovery-config.yaml schedule.feedback_ingestion
• Manual -- human requests "Ingest customer feedback for [product]"
• Event-driven -- triggered by high-volume feedback spikes (if monitoring configured)

Prerequisites

• discovery-config.yaml exists with products[*].user_channels configured
• .factory/discovery/feedback-state.yaml exists (or will be created on first run)
• research-agent is the canonical MCP caller (DF-002); this skill spawns it for review searches (perplexity_search for public app/review-site lookups)
• GitHub API and Slack API access available for channel ingestion

Supported Channels

Channel	Access Method	What It Captures
GitHub Issues	gh CLI / GitHub API	Feature requests, bug reports, discussions
GitHub Discussions	gh CLI / GitHub API	Community questions, feature ideas
Slack/Discord	Webhook archive or API	Customer messages in feedback channels
App Store Reviews	perplexity_search (public)	iOS/Android app reviews
G2/Capterra Reviews	perplexity_search (public)	Enterprise software reviews
Support Tickets	API integration (configurable)	Customer support issues
NPS/Survey Results	File import (.factory/surveys/*.csv)	Structured survey responses

Channel Configuration

Channels are configured per-product in discovery-config.yaml:

products:
  - name: "my-product"
    user_channels:
      - type: "github-issues"
        repo: "org/product"
        labels_feature: ["feature-request", "enhancement"]
        labels_bug: ["bug", "defect"]
        since: "last-ingestion"

      - type: "github-discussions"
        repo: "org/product"
        categories: ["Ideas", "Feature Requests"]
        since: "last-ingestion"

      - type: "slack-channel"
        workspace: "company"
        channel: "#product-feedback"
        since: "last-ingestion"

      - type: "app-reviews"
        search_terms: ["my-product"]
        platforms: ["ios", "android"]
        since: "30d"

      - type: "review-sites"
        search_terms: ["my-product"]
        sites: ["g2.com", "capterra.com"]
        since: "30d"

      - type: "survey-import"
        path: ".factory/surveys/"
        format: "csv"

Ingestion Workflow

Step 1: Read Channel Config

Read discovery-config.yaml and load the user_channels array for the target product. If no channels are configured, exit with no-op.

Step 2: Load Ingestion State

Read .factory/discovery/feedback-state.yaml for last ingestion timestamps per channel. If the file does not exist (first run), initialize with empty state.

# .factory/discovery/feedback-state.yaml
last_run: YYYY-MM-DDTHH:MM:SSZ
channels:
  github-issues:
    last_ingested: YYYY-MM-DDTHH:MM:SSZ
    items_total: N
  github-discussions:
    last_ingested: YYYY-MM-DDTHH:MM:SSZ
    items_total: N
  # ... per channel

Step 3: Fetch New Items Per Channel

For each configured channel:

GitHub Issues:

gh issue list --repo org/product --label feature-request --state open \
  --json number,title,body,labels,createdAt,author,comments \
  --limit 100

Filter to items created/updated since last_ingested.

GitHub Discussions:

gh api graphql -f query='{ repository(owner:"org", name:"product") {
  discussions(first:50, categoryId:"...") { nodes { title body createdAt } }
}}'

Slack/Discord:

• Via the Slack/Discord API: fetch messages from configured channel since last ingestion
• Parse for feature requests, complaints, praise

App Store Reviews:

• Via research-agent -> perplexity_search: "[product name] app reviews [platform] [month year]"
• Extract individual reviews with ratings and text

Review Sites (G2/Capterra):

• Via research-agent -> perplexity_search: "[product name] reviews [site] [month year]"
• Extract review summaries with ratings

Survey Import:

• Read CSV files from .factory/surveys/ directory
• Expected columns: date, respondent_id, type (NPS/CSAT/custom), score, comment
• Process new files not previously ingested

Step 4: Categorize Each Item

Assign each ingested item to one category:

Category	Signal Words / Patterns	Priority
Feature Request	"would be great if", "please add", "feature request", enhancement labels	HIGH
Bug Report	"broken", "doesn't work", "error", "crash", bug labels	HIGH
Pain Point	"frustrated", "difficult", "can't figure out", "workaround"	MEDIUM
Praise	"love", "great", "amazing", "thank you", 4-5 star reviews	LOW
Question	"how do I", "is it possible", "documentation", "?"	MEDIUM

Categorization uses the business-analyst agent's judgment, not just keyword matching. Context and tone matter.

Step 5: Deduplicate

Cluster similar items to avoid inflated counts:

1. Group items by category
2. Within each category, compute semantic similarity
3. Cluster items with >0.80 similarity into a single entry
4. Track count per cluster (frequency signal)
5. Preserve the most detailed/articulate item as the cluster representative

Step 6: Produce Feedback Digest

Write .factory/discovery/feedback-digest-YYYY-MM-DD.md:

---
document_type: feedback-digest
date: YYYY-MM-DD
product: [product-name]
channels_ingested: [list]
total_items: [N]
new_since_last: [N]
---

# Feedback Digest: YYYY-MM-DD

## Summary
| Category | Count | Trend vs Last Period |
|----------|-------|---------------------|
| Feature Requests | [N] | up/down/flat |
| Bug Reports | [N] | up/down/flat |
| Pain Points | [N] | up/down/flat |
| Praise | [N] | up/down/flat |
| Questions | [N] | up/down/flat |

## Top Feature Requests (by frequency)
| Rank | Request | Sources | Count | First Seen | Trend |
|------|---------|---------|-------|-----------|-------|
| 1 | [request] | GH, Slack, G2 | [N] | [date] | up |

## Top Pain Points
| Rank | Pain | Sources | Count | Severity |
|------|------|---------|-------|---------|
| 1 | [pain] | [sources] | [N] | HIGH/MED/LOW |

## Emerging Signals (new this period)
[Items not seen in any previous digest]

## Raw Items
[Appendix: all ingested items with source, date, category, text]

Step 7: Update State

Update .factory/discovery/feedback-state.yaml with:

• Current timestamp per channel
• Cumulative item counts
• Run metadata

Quality Gate

• All configured channels scanned (failures logged, not blocking)
• Feedback items categorized with rationale (not keyword-only)
• Deduplication applied -- similar items clustered, no inflated counts
• Feedback digest written to .factory/discovery/feedback-digest-YYYY-MM-DD.md

Quality Criteria

• All configured channels attempted (failures logged, not blocking)
• Items categorized with rationale (not just keyword matching)
• Deduplication clusters similar items (no inflated counts)
• Frequency counts are accurate (cluster size, not raw count)
• Trends compare against previous digest (if available)
• State file updated (no re-processing on next run)
• Raw items preserved in appendix for traceability