Support Triage Agent Template

A read-only, evidence-graded support investigation workflow. The agent turns messy customer tickets into structured triage reports with confidence scoring, known-issue search, and a customer-ready response draft.

This file is the entry point for Claude Code. The companion AGENTS.md carries the same workflow rules in a portable form for other AGENTS-aware clients (Codex, GitHub Copilot CLI, Cursor).

How to run

The shipped slash command is /triage:

/triage <paste ticket text or path to a ticket file>

The command delegates to the read-only agent in agents/support-triage-agent.md and runs the full workflow described there: intake → parallel research → known-issue search → synthesis → review.

Project structure

• agents/support-triage-agent.md — agent definition (workflow, rules, read-only tool allow-list, anti-hallucination guards); symlinked into .claude/agents/ so the repo also works as a plain Claude Code project
• commands/triage.md — /triage slash command; symlinked into .claude/commands/
• .claude/settings.json — read-only allow/deny defaults (no destructive verbs, no flag/ticket mutations)
• .mcp.json.example — connector cookbook for common support stacks (GitHub, Sentry, Linear, Slack, DeepWiki, Context7); copy to .mcp.json and fill in
• skills/ — composable workflow skills (intake, known-issue search, triage report, response drafting, escalation, redaction, log evidence, reproduction steps, KB article)
• skills/*/references/ — canonical report shapes (triage report, customer response, escalation brief), shipped inside the skills so installed plugins can find them
• examples/tickets/ and examples/mock-sources/ — synthetic tickets and an offline mock bug tracker, docs, release notes and status page
• reports/ and docs/demo/ — sample output and recorded E2E runs
• .claude-plugin/plugin.json (Claude Code) and plugin.json (GitHub Copilot CLI) — plugin manifests sharing the same skills, agent and command
• docs/architecture.md — mermaid view of the workflow
• docs/connectors.md — MCP connector cookbook with copy-paste examples
• AGENTS.md — same rules, portable form for other clients

What this template is not

• It is not an integration with any specific company. The sample domain is synthetic.
• It does not auto-respond to customers. Every report is a draft for a human support engineer to review.
• It does not write to production systems. All investigation is read-only.

Adapting it

To use this on a real product, edit four places:

1. skills/ticket-intake/SKILL.md — replace generic product areas with yours.
2. .mcp.json.example → .mcp.json — wire your real MCP servers (your bug tracker, your error monitor, your docs source).
3. Add product-specific diagnosis skills under skills/<your-product-area>-diagnosis/.
4. examples/tickets/ and reports/sample-triage-report.md — replace with sanitized real examples.

Keep the report shape stable so reports stay comparable across tickets and across investigators.

Data governance reminder

When MCP servers connect to real customer systems, customer data passes through the LLM provider. Before pointing this template at production:

• Confirm your LLM provider has a DPA covering customer data.
• Prefer Zero Data Retention or in-boundary routing (Bedrock / Vertex / equivalent).
• Scope MCP API keys per engineer with audit logging.
• Consider running with MCP project tools removed for low-trust contexts (demos, evals); the agent degrades gracefully and asks the customer to verify what it cannot check.