When I investigate an issue, the evidence rarely lives in one place. There are Jira tickets, Slack conversations, client feedback, and dashboards. Each tells only part of the story.
Bringing those pieces together takes time, especially when I’m trying to understand how widespread a problem is and whether it needs to be prioritized. I wanted to see if Claude Code could help.
Drawing inspiration from Product School and Enterpret’s Feedback Intelligence plugin, I built my own plugin and adapted it to the tools and workflows I use.
I’d already experimented with gathering scattered information for my weekly Slack update. This time, I wanted to answer a different question: how much confidence could I place in the patterns and conclusions Claude identified?
I designed the plugin to give me specific things to check before relying on its output.
How the plugin works
The plugin has three commands: /escalation-triage, /bug-triage, and /prd-evidence. Together, they help me spot emerging problems, investigate related reports, and gather information for prioritization.
To show how the commands work together, I’ve used a fictional scenario involving a refund that times out and gets retried, resulting in a duplicate posting.
The outputs below are fictional illustrations based on the plugin’s instructions, not results from actual Claude Code runs. The reports and numbers are invented.
1. Spot emerging escalations
The /escalation-triage command looks for emerging escalations, priority changes, and spikes in reported problems.
Command:
/escalation-triage "refunds" --hours 24
Illustrative output (excerpt):
Search scope: Last 24 hours. Jira keyword search and two Slack channels: #incidents and #payments-alerts.
Severity Finding Source Affected P0 Refunds posting twice after a retry Jira + Slack 2 accounts P1 Partial refunds failing on discounted orders Slack 4 orders P2 Spike in refund-status tickets Jira 1 reporting team
P0: BILL-412 was raised from Medium to Highest overnight. Slack reports describe the same symptom.Limitations: Only two Slack channels were scanned. Private channels and other discussions weren’t covered.
The duplicate-refund reports stand out as the most urgent finding. The command also surfaces a separate problem reported only in Slack, which a Jira-only search would miss.
2. Investigate related reports
The /bug-triage command groups reports into clusters, assesses severity, and estimates how many distinct accounts, orders, or teams are affected.
Command:
/bug-triage "refunds"
Before searching, Claude asks whether I want to focus on a particular Slack channel. In this example, I don’t specify one, so it searches the public channels available to it.
Illustrative output (excerpt):
Cluster: Refund posted twice to the same account
Severity: P0
Affected: 2 distinct accounts, based on 2 Billing tickets found in the 30-day investigation.
Description: A refund retried after a timeout posted a second time. The ledger no longer matches the payment processor.
Reproduction HYPOTHESIS: The retry may not check whether the first attempt already succeeded.
Recommended next step: Engineering to confirm the cause. Operations to correct the affected accounts.
Dashboard cross-check: 5 unmatched refund records in the last 30 days, reported separately from the account count.
Limitations: Search covers accessible Jira projects and public Slack channels. Monitoring-bot tickets and reports without enough detail to cluster were excluded.
The output identifies a possible cause without treating it as confirmed. It also reports the dashboard findings separately from the Jira account count because the records may overlap.
3. Gather evidence for prioritization
The /prd-evidence command pulls together the problem statement, scale, severity, supporting reports, possible impact, and limitations. It also prepares material for a product requirements document (PRD) or stakeholder discussion.
Command:
/prd-evidence "refund retry double-posting"
Claude first asks which client-feedback view and dashboard to use. In this example, I select Support feedback and the client-conversation dashboard.
Illustrative output (excerpt):
Problem: Operations can’t safely retry a refund that times out. Some refunds are posted twice, leaving the ledger out of balance with the payment processor.
Scale:
- Jira: 9 tickets representing 6 distinct accounts over 90 days.
- Trend: 1 affected account in July, 1 in August, and 4 in September.
- Client-feedback tool: 18 entries in the duplicate-refund group.
- Client-conversation dashboard: 25 conversations about duplicate refunds.
Impact HYPOTHESIS: Each duplicate may require manual correction, reconciliation work, and a support conversation.Limitations:
- Feedback entries and dashboard conversations may represent some of the same interactions. Their counts shouldn’t be added.
- Two Jira tickets don’t establish whether a retry occurred.
- Slack coverage is limited to accessible public channels.
- Processor and ledger records need to be compared to confirm the cause.
Suggested stakeholder summary: Refund retries can post twice, leaving the books out of balance. The number of affected accounts increased from one in July to four in September.
The first command flags something that needs attention, the second organizes related reports and possible causes, and the third assembles information for a prioritization decision.
Having that information in one place doesn’t automatically make the conclusions reliable. That’s why I built additional checks into the instructions.
Count the impact, not just the reports
In the refund example, nine Jira tickets represent six distinct accounts. Some accounts have multiple tickets about the same problem. That’s why I instructed Claude to count distinct affected accounts rather than assume each ticket represents a different account.
I also wanted to know whether reports were concentrated within a particular team. In the separate issue involving refund statuses stuck in Pending, Operations submitted 11 out of 14 reports. I’d want to understand why. Is that team seeing something others aren’t, or is the problem more widespread than the reports suggest?
The plugin flags that concentration so I can investigate it before drawing conclusions about the scale of the problem.
Check whether the sources overlap
In the refund example, the client-feedback tool contains 18 entries about duplicate refunds, while the dashboard shows 25 conversations about the same problem.
But these may not represent separate incidents. For example, a client might contact Customer Support about a duplicate refund. That conversation could appear in the dashboard and also be recorded in the client-feedback tool.
Adding the two counts would give me 43 records, but not necessarily 43 distinct complaints or affected clients.
That’s why I instructed Claude to report the counts separately and flag possible overlap. The plugin doesn’t identify which individual records are duplicates, so I’d still need to compare them before using the numbers to assess how widespread the problem is.
Make uncertainty visible
In the refund example, Claude suggests that a retry may have posted a refund twice because the system didn’t check whether the first attempt succeeded. But the available reports don’t confirm that explanation. I’d need to examine the payment processor and ledger records to establish what happened.
There’s also the question of whether Claude has found all the relevant information. Its Slack search is limited to the channels it can access, so important conversations could be missing.
I wanted the plugin to make both limitations clear: what the available evidence doesn’t yet establish, and what information might be missing from the investigation. That’s why I instructed Claude to label possible causes and impact estimates as hypotheses, identify which sources it searched, and flag anything that still needs verification.
The judgment still sits with me
Making those limitations visible helps me evaluate the evidence, but it doesn’t tell me what decision to make. The plugin can gather reports, identify patterns, and suggest possible explanations. It can’t determine whether the evidence is strong enough to justify prioritizing a fix.
That’s still my responsibility. I need to understand what the sources actually show, where the conclusions rely on assumptions, and what needs further investigation before making a recommendation.