Best Practices Playbook

This playbook is a practical guide to using ADM effectively across roles—data engineers, analytics engineers, data stewards, and business stakeholders.

It focuses on:

• Asking better questions (prompting)

• Choosing modes intentionally (Conversation vs Workflows vs Notebooks)

• Building repeatable operations (triage workflows)

• Collaborating safely (multi-user + permissions)

• Grounding answers using Knowledge Base and MCP

How to Access These Features

• Conversation: left nav → Conversation / New Conversation

• Workflows: left nav → Understanding Workflows / Workflows

• Agents: left nav → Understanding Agents

• Business Notebooks: left nav → Business Notebooks

• Knowledge Base: left nav → Knowledge Base (upload docs)

• Collaboration: open a conversation → Share / Add participants

• MCP: left nav → Understanding MCP Server (and configured integrations)

Playbook by role

Data Engineer (Triage + Remediation)

Best outcomes when you:

• Ask for root cause hypotheses + evidence

• Compare “before vs after”

• Request “what changed” upstream

• Standardize triage into workflows

Go-to prompts

• “Compare today vs yesterday for this asset: what changed in quality, freshness, schema?”

• “List the top 3 likely causes and the fastest checks to confirm.”

Analytics Engineer (Stability + Correctness)

Best outcomes when you:

• Ask for rule-level breakdowns

• Validate assumptions about transformations

• Request impact analysis (downstream assets/dashboards)

Go-to prompts

• “Which downstream dashboards depend on this asset and are impacted?”

• “Which rule failures are most likely due to transformation logic?”

Data Steward (Governance + Standards)

Best outcomes when you:

• Reference documented policies and definitions (Knowledge Base)

• Ask for “standard vs exception”

• Request stakeholder-ready summaries

Go-to prompts

• “According to our standards, what should quality thresholds be for this dataset?”

• “Summarize non-compliant assets and recommend which policies to add.”

Business User (Impact + Decisions)

Best outcomes when you:

• ask for business impact first

• request non-technical language

• ask for “what should I do now”

Go-to prompts

• “Is this metric safe to use today? If not, what’s the recommended workaround?”

• “Summarize impact on revenue/finance reporting in 8 bullets.”

Operational Best Practices (What High-performing Teams Do)

1) Start with a “triage checklist” Approach

Ask ADM to answer, in order:

1. What happened (symptoms)?

2. When did it start?

3. What changed?

4. What is impacted?

5. What do we do next?

This avoids skipping to conclusions.

2) Turn Repeatable Work into Workflows

Once you run the same analysis more than twice, convert it into a workflow:

• Consistent steps

• Consistent output format

• Reusable across teams

3) Use the Knowledge Base for “single source of truth”

Upload:

• Governance docs

• SLAs

• Definitions of key metrics

• Runbooks

• Incident playbooks

Then ask ADM: “Use Knowledge Base sources and cite them.”

4) Use Collaboration to Reduce Decision Latency

Bring in:

• Data platform (pipelines)

• Domain owners (definitions)

• Consumers (impact confirmation)

Use @mentions + short asks:

• “Can you confirm whether this is expected seasonality?”

• “Is the SLA for this report 9 AM or 11 AM?”

5) Always Ask for “next checks”

For investigations, require ADM to include:

• 3–5 verification checks

• the fastest path to confirm root cause

• recommended owner team

Real-world Scenarios (Templates You Can Copy)

Scenario A — Freshness Breach on a Tier-1 Table

Prompt: “Investigate why daily_revenue_summary is delayed today. Compare to prior runs, identify the likely cause, list impacted downstream assets, and provide a recommended mitigation plan.”

Expected output:

• Summary (5 bullets)

• Likely root cause + evidence

• Impacted dashboards/reports

• Next steps + owners

Scenario B — Schema Drift Broke a Pipeline

Prompt: “Schema drift detected on orders. Identify added/removed/modified columns, assess likely pipeline breakpoints, and propose a remediation plan.”

Expected output:

• Changes table (column, change type, risk)

• Suspected breaking dependencies

• Recommended fix steps

Scenario C — Data Quality Score Dropped Suddenly

Prompt: “Quality score for customer_master dropped from 97% to 85% in 24 hours. Show failing rules, affected row counts, and what upstream change could explain it. Provide verification steps.”

Expected output:

• Failing rules table

• Pattern analysis

• Verification checks

• Prevention recommendation

Scenario D — Business Asks “Can I trust the dashboard?”

Prompt: “For the executive revenue dashboard, confirm whether data is reliable today. If not, explain impact and recommended decision guidance in plain language.”

Expected output:

• Trust status (green/yellow/red)

• What is impacted

• Recommended actions / workarounds

• Links to evidence (where supported)

Common Pitfalls and Fixes

Pitfall: Unclear Ownership

Fix: Ask ADM to suggest owner teams based on domain/asset tags (or your org’s structure).

Pitfall: Too Much Detail for Stakeholders

Fix: Ask for “exec summary + appendix,” where appendix contains technical details.

Pitfall: Too Many One-off Conversations

Fix: Create a workflow + notebook template for recurring reporting.

Checklist: “Good ADM request”

• Asset(s) named (tables, domains, dashboards)

• Time window defined

• Goal stated (investigate / summarize / decide)

• Output format specified (bullets/table)

• Ask for evidence/citations where appropriate

• Ask for next steps + verification checks