Group duplicate customer requests without losing their sources
Turn scattered support requests into problem themes linked back to the original customer evidence.
Target user
Product managers receiving feature requests through support, email, and feedback forms.
Trigger
Backlog refinement, or a batch of new requests arrives and similar ideas are hard to spot.
Why run it repeatedly
Run during backlog refinement or when request volume makes manual deduplication disruptive; measure actual volume with the pilot team.
What you receive
A proposed set of problem themes, candidate duplicates, original request links, customer impact evidence, and unanswered questions.
Pain evidence
Direct first-person reports describe duplicate requests arriving through email/forms and spreadsheet or backlog cleanup. Frequency and willingness to change existing feedback tooling remain unvalidated.
Alternatives and remaining gap
Productboard and similar products provide feedback portals, duplicate linking, and roadmap views. A deterministic classifier/script may be enough if records are already normalized; AI helps explain semantic similarities and preserve nuance, but should not decide merges.
Execution flow
- Search a defined time window or supplied conversation batch and fetch only relevant requests.
- Normalize each request into the underlying user problem; preserve the original wording and URL.
- Compare candidate themes against existing Linear issues and group only similar problems, never just identical feature wording.
- List support counts only for records actually retrieved; surface counterexamples and uncertain matches.
- Draft a source-linked evidence brief and wait for product review before creating or updating backlog records.
Required tools
- Intercom: search_conversations, get_conversation, search_contacts
- Linear: list_issues, get_issue, create_issue (after approval)
- Notion: notion-search, notion-fetch, notion-create-pages (after approval)
Prerequisites and costs
Intercom conversation read scope; Linear issue read scope and write scope only if approved; a Notion workspace shared with the client if publishing the review brief. Actual app permissions and plans apply.
Run this prompt
Review these customer requests from [date range or conversation IDs]. Find repeated underlying problems, not merely similar feature wording. Preserve every source conversation link and customer identifier only when permitted. Search Linear for existing themes. Show uncertain groupings and counterexamples. Do not create, merge, or update records; return a draft evidence brief for review.
Review checkpoints
Review theme boundaries, customer privacy, evidence counts, and existing issue links. A human decides whether requests are genuinely duplicates and whether to update the roadmap.
Limitations
A small Reddit thread is evidence that the task occurs, not proof of prevalence. Productboard and dedicated feedback systems may already solve intake and deduplication. MCP adds value only where the source systems remain fragmented and evidence must be synthesized.
Manual work and missing capabilities
Map request channels to Intercom, redact PII, define the problem taxonomy, and have a PM approve each proposed cluster.
Failure handling
Use bounded retries for read failures and mark partial results incomplete. Keep source record IDs and links. Do not guess across missing or conflicting identifiers. Search for an existing target record before any approved write, then retrieve it to verify.
Pilot acceptance criteria
On a reviewed sample, every group has source links; known false-positive pairs remain separate; all records are accounted for; no automatic merge or roadmap commitment occurs.
Product fit
MCPfolio recipe for read-only theme discovery. Consider a workflow package only if teams repeatedly need ingestion, deduplication history, and a durable review queue.
Demo storyboard
Show three synthetic requests with source links, a false-positive pair kept separate, and one candidate theme linked to an existing Linear issue.
Official sources
Pain reported by users; official tool docs/catalog checked; workflow not runtime tested
