What problem is worth solving?
Teaching scenario: a complaint mentions delay, damaged packaging and a poor response, but the system permits one category. Preserve issue relationships and actionable ownership.
Reduce missed issues and connect customer feedback to product, fulfillment and service improvements.
- WORKING ROLES
- Customer experience, fulfillment, product quality and FDE
How does an FDE turn symptoms into a solvable problem?
The following discovery and implementation exercise uses a synthetic scenario.
Read issue units
Split ten invented complaints into issue units with ticket IDs and source locations; compare omissions against single labels.
Inspect the counting unit
Three contacts and three issues in one message are different dimensions. Define customers, tickets and issues before aggregation.
How does the solution enter the workflow?
- 01
De-identify
Remove identities and contact details; split factual issue spans.
- 02
Group
Use stable categories and preserve emerging unknown themes.
- 03
Review
Sample source spans for omissions, duplicates and misclassification.
- 04
Act
Assign a testable question, owner and review window to each theme.
Which decisions make the solution reliable?
Clusters are leads
Damage frequency may reflect carriers, packaging or sampling; check batch and fulfillment evidence before assigning cause.
Keep emerging themes
Review unstable categories; rebase comparisons when definitions change.
One ticket, three issues, one customer
Invented complaint: “Two days late, crushed box, and support gave no explanation.”
- Split
- T-1/I-1 delay; I-2 packaging; I-3 response.
- Count
- Three issues, one ticket, one customer.
- Action
- Check shipment time, packaging batch and response, with appropriate owners.
What should you do and deliver?
Define complaint issue units
- OUTPUT
- Discovery + coding guide
- PASS CONDITION
- Each issue has a source span
Deduplicate and count dimensions
- OUTPUT
- Issue table + aggregation
- PASS CONDITION
- Customers and issues are not conflated
Test omissions and theme drift
- OUTPUT
- Labels + review report
- PASS CONDITION
- New themes are not forced into old labels
How do you prove completion and reuse the work?
Acceptance checks
Assess omissions, classification, duplicate counting and ownership; keep causes testable.
Reusable assets
Reuse issue units, source spans and multidimensional counts for reviews, interviews and support, with business-owned taxonomies.
Further reading and original sources
- Building effective agents ↗Anthropic · 2024-12-19
- Demystifying evals for AI agents ↗Anthropic · 2026-01-09