Feedback to backlog

When to run

• Recurring: pre-grooming weekly sweep (recommended)
• "Consolidate the export-API requests" / roadmap planning season

Computer scopes

• Support + CS + sales: tickets, call notes, meeting summaries carrying requests
• Community: synced forum/board threads
• Product tracker: existing backlog items (dedupe target, draft destination)
• CRM: account tier and ARR for demand weighting

Workflow

1. Extract request-bearing passages from the window's records; normalize to the underlying capability being asked for, not the phrasing.
2. Dedupe against existing backlog items first — the most common failure is a fourth duplicate of a tracked item. Attach new evidence to existing items rather than creating twins.
3. For genuinely new requests, cluster and consolidate: one draft item per capability with problem statement in customer language, distinct-account count, ARR-weighted demand, and 2–3 cited verbatims.
4. Rank by demand weight × strategic tags (from the product brief) and mark conflicts where requests pull opposite directions — surfacing tension is the job, resolving it is product's.
5. Queue drafts and evidence-attachments for PM approval; on approval, write to the tracker with all links intact.

Guardrails

• Autonomy: L1. Tracker writes are drafted and human-approved; evidence attachment to existing items may be promoted to L2 after four clean runs.
• Counts are distinct-account and auditable; no double-counting a loud champion.
• Verbatims exact and tier-attributed, not name-attributed, in shared views.

Output

Eval fixtures

• Set: 300 sandbox records containing 14 request clusters, 6 of which duplicate existing backlog items; seeded ARR weights.
• Rubric: dedupe correctness (the 6 must attach, not duplicate); cluster completeness; count and weight accuracy; verbatim fidelity.
• Ship gate: duplicate-creation rate 0; cluster recall ≥ 0.85.