| # Feedback Clusterer |
|
|
| ## Objective |
|
|
| Turn noisy feedback streams into a small set of actionable themes with evidence, owners, and suggested next steps. |
|
|
| ## Trigger |
|
|
| - Schedule: daily or weekly. |
| - Event: launch, incident, release, community spike, support backlog growth, or high-volume issue labels. |
| - Manual bootstrap/debug command: "cluster feedback from the last week." |
|
|
| ## Intake |
|
|
| - GitHub issues, PR comments, Linear tickets, Slack threads, support tickets, app reviews, forum posts, social posts, and survey responses. |
| - Existing product areas, labels, owners, roadmap themes, and known duplicates. |
|
|
| ## Agents |
|
|
| - Collector: gathers feedback and preserves source links. |
| - Clusterer: groups items by user problem, not by surface wording. |
| - Skeptic: checks whether clusters are overgeneralized or missing counterexamples. |
| - Reporter: produces themes, examples, confidence, and next actions. |
|
|
| ## Workspace And Permissions |
|
|
| - Prefer read-only access to feedback systems. |
| - Allow creating summary docs, issues, or labels when permitted. |
| - Require human approval for closing user reports, changing priorities, or sending customer-facing replies. |
|
|
| ## Durable State |
|
|
| - Query windows, source filters, processed item IDs, cluster names, representative links, confidence, and follow-up owners. |
|
|
| ## Loop Steps |
|
|
| 1. Define the time window and source filters. |
| 1. Collect raw items with stable links. |
| 1. Delegate collection, clustering, skepticism, and reporting to separate roles when source volume is high. |
| 1. Deduplicate exact repeats and known duplicates. |
| 1. Cluster by underlying user need, failure mode, or request. |
| 1. Attach representative examples and counterexamples. |
| 1. Rank by frequency, severity, strategic relevance, and confidence. |
| 1. Produce action items or escalate ambiguous themes. |
|
|
| ## Verification Gates |
|
|
| - Every cluster has source links. |
| - High-impact claims include representative examples. |
| - The report distinguishes frequency from severity. |
| - The loop records processed IDs to avoid double counting. |
|
|
| ## Budget And Exit |
|
|
| - Max runtime: 30-90 minutes depending on source volume. |
| - Max items: set a sampling cap before the run. |
| - Stop when clusters are stable, source rate limits block collection, or themes require product judgment. |
|
|
| ## Escalation |
|
|
| Escalate when feedback includes security issues, legal concerns, customer-specific commitments, sensitive data, or prioritization conflicts. |
|
|
| ## Loop Instruction |
|
|
| ```text |
| Cluster feedback from <sources> for <time window>. |
| Group by underlying user problem, keep source links, separate frequency from severity, and record processed IDs. |
| Produce the top themes with representative examples, confidence, suggested owners, and next actions. |
| Do not close issues or send external replies without approval. |
| ``` |
|
|
| Example automation: run daily or weekly, or trigger after a launch, incident, release, support backlog spike, or high-volume issue label. |
|
|
| ## Failure Modes |
|
|
| - The loop clusters by keywords instead of user problems. |
| - High-volume low-severity noise hides rare severe issues. |
| - The report loses links back to evidence. |
| - The same item is counted repeatedly across mirrored systems. |
|
|
| ## References |
|
|
| - [Codex Loops: What Boris Cherny Gets Right About Managing Agent Work](https://www.developersdigest.tech/blog/codex-loops-boris-cherny-agent-routines) - Includes feedback clustering as a recurring agent loop pattern. |
| - [How we built our multi-agent research system](https://www.anthropic.com/engineering/multi-agent-research-system) - Shows orchestrator-worker patterns for collecting, checking, and reporting evidence. |
|
|
