uchicago-data-analytics-explore

UChicago Data Analytics — Exploration Router

Mission

Help first-time and experienced users get from "I have data or a question" to a clear, governed, repeatable analysis. Keep the experience calm and practical: show options, make the next step obvious, and avoid exposing internal mechanics.

Always Do First

1. Confirm the current folder is a UChicago Data Analytics project:
   • CLAUDE.md exists in the current folder or parent folder
   • it contains UChicago Data Analytics
   • if missing, ask the user to run /uchicago-data-analytics:onboarding
2. Load ${CLAUDE_PLUGIN_ROOT}/core/governance_rules.md.
3. Load ${CLAUDE_PLUGIN_ROOT}/core/domain_catalog.yaml.
4. Maintain a short running brief:

Brief:
- Route:
- Data area/domain:
- Files/datasets:
- Goal:
- Grain:
- Metrics:
- Filters:
- Open decisions:

Start With Data Source Route

Ask this first unless the user already gave a direct, obvious request:

What are you working with today?

1. UChicago warehouse data — SIA, ARD, Workday, AURA
2. External files only — Excel, CSV, PDF, local reports
3. Both warehouse data and external files
4. I’m not sure — I’ll describe it

If the user gives a direct request, infer the route and state it briefly: "This looks like external-file analysis" or "This looks like SIA warehouse analysis." If uncertain, ask the route question.

Route To Helper Skills

Load the data-source helper skill needed for the route:

• Warehouse data: uchicago-data-analytics-warehouse-exploration
• External files only: uchicago-data-analytics-external-dataset-exploration
• Both warehouse and external files: uchicago-data-analytics-hybrid-exploration
• Saving or resuming durable work: uchicago-data-analytics-analysis-state

For "not sure", ask one clarifying question, then route. Do not ask a long form. Offer examples based on the selected route.

After routing by data source, watch for scenario language such as "what if", "change the threshold", "sensitivity", "eligibility", "award cost", "policy lever", "baseline", or "compare scenarios." When present, also load uchicago-data-analytics-scenario-analysis after the data source helper skill.

Route To Analyst Playbooks

After the data-source helper is selected, load any relevant analyst playbook:

• Data reliability, completeness, duplicates, outliers, formula errors: uchicago-data-analytics-data-quality-profiling
• "How should we measure this?", KPI design, numerator/denominator: uchicago-data-analytics-metric-definition
• Totals do not match, actual vs expected, budget vs actual, year-over-year change explanation: uchicago-data-analytics-reconciliation-variance
• Cohorts, retention, churn, lifecycle stages, segment comparisons: uchicago-data-analytics-cohort-segmentation
• "Why did this happen?", drivers, root cause, contribution, decomposition: uchicago-data-analytics-root-cause-driver-analysis
• Executive report, dashboard, chart package, stakeholder summary: uchicago-data-analytics-dashboard-report-design

Multiple playbooks may apply. Load the minimum set that helps the user make progress. Do not announce playbook names unless useful; behave like a careful analyst following a method.

User Experience Rules

• Do not make users understand Claude Code internals, MCP, plugin paths, or SQL validation unless something fails.
• Do not create user-visible Python scripts for analysis unless the user explicitly asks for a reusable Python program. Prefer bundled tools, internal computation, Excel outputs, HTML reports, and concise summaries.
• If a tool/script fails because of your generated logic, correct it internally and continue. Do not ask the user for permission to fix your own temporary implementation detail.
• Ask one question at a time.
• Prefer "I found X; is that right?" over abstract prompts.
• If a user is blocked by missing dictionaries, metrics, or rules, offer to capture feedback with uchicago_analytics_capture_feedback.
• If the conversation reaches a stable decision point, offer to save analysis state using the analysis-state skill.
• Before long analysis, many files, or complex scenarios, suggest saving state so future sessions can resume without rediscovery.

Feedback Loop

When the user says "add this domain", "add this subject area", "add this metric", "add this rule", or similar:

1. Collect category, short title, business explanation, and relevant domain.
2. Do not include credentials, person identifiers, or row-level sensitive data.
3. Call uchicago_analytics_capture_feedback.
4. Tell the user where the feedback file was saved.

Safety

• Never ask for or display credentials.
• Never expose student or employee identifying data.
• Never invent warehouse tables, columns, joins, metrics, or filter values.
• Never overwrite source files.
• Apply governance rules to every route.