--- name: Analyst description: Extract insights from data with SQL, visualization, and clear communication of findings. metadata: {"clawdbot":{"emoji":"🔍","os":["linux","darwin","win32"]}} ---

Data Analysis Rules

Framing Questions

• Clarify the decision being made — analysis without action is trivia
• "What would change your mind?" surfaces the real question
• Scope before diving in — infinite data, limited time
• Hypothesis first, then test — fishing expeditions waste time

Data Quality

• Validate data before analyzing — garbage in, garbage out
• Check row counts, date ranges, null rates first
• Duplicates hide in joins — always verify uniqueness
• Source definitions matter — revenue means different things to different teams
• Document assumptions — future you needs context

SQL Patterns

• CTEs over nested subqueries — readable beats clever
• Aggregate before joining when possible — performance matters
• Window functions for running totals, ranks, comparisons
• CASE statements for categorization — clean logic
• Comment non-obvious filters — why are we excluding these?

Analysis Approach

• Start with the simplest cut — don't overcomplicate early
• Cohorts reveal what aggregates hide — when did users join?
• Time series need seasonality awareness — don't compare Dec to Jan
• Segmentation surfaces patterns — average obscures variation
• Correlation isn't causation — but it's where to look

Visualization

• Chart type matches data: trends (line), comparison (bar), distribution (histogram)
• One message per chart — don't overload
• Label axes, title clearly — standalone comprehension
• Color with purpose — highlight, don't decorate
• Tables for precision, charts for patterns

Communicating Findings

• Lead with the insight, not the methodology
• So what? Now what? — always answer these
• Confidence levels matter — don't oversell noisy data
• Recommendations are opinions — label them as such
• Executive summary first, details available — respect their time

Stakeholder Relationship

• Understand their mental model before presenting
• Regular check-ins prevent surprise requests
• Push back on bad questions — help them ask better ones
• Data literacy varies — adjust explanation depth
• Their intuition is data too — triangulate

Tools

• Right tool for the job: SQL for querying, spreadsheets for ad-hoc, BI for dashboards
• Reproducibility matters — scripts over clicking
• Version control analysis code — changes need history
• Automate recurring reports — manual refresh doesn't scale

Common Mistakes

• Answering the wrong question precisely
• Cherry-picking data that confirms expectations
• Overfitting: explaining noise as signal
• Death by dashboard: metrics nobody checks
• Analysis paralysis: perfect insight never delivered