Exploratory Analysis Process

Overview

This skill guides you through systematic exploration of unfamiliar datasets when you don't yet have a specific question to answer. Unlike hypothesis-testing (where you test a specific claim) or guided-investigation (where you answer a specific question), exploratory analysis helps you discover what's interesting in the data and identify what questions you should be asking.

Exploratory analysis is appropriate when:

• You have a new dataset and need to understand what's in it
• The user says "Just see what's interesting" or "Tell me what stands out"
• You need to discover patterns before formulating specific questions
• You're looking for unexpected insights or anomalies
• You want to generate hypotheses for later testing

Prerequisites

Before using this skill, you MUST:

1. Have data imported into SQLite database using the importing-data skill
2. Have data quality validated and cleaned using the cleaning-data skill (MANDATORY - never skip)
3. Have created an analysis workspace (just start-analysis exploratory-analysis <name>)
4. Have NO preconceived hypotheses or specific questions (if you do, use hypothesis-testing or guided-investigation instead)
5. Be familiar with the component skills:
   • understanding-data - for data profiling
   • writing-queries - for SQL query construction
   • interpreting-results - for result analysis
   • creating-visualizations - for text-based visualizations

Mandatory Process Structure

You MUST use TodoWrite to track progress through all 5 phases. Create todos at the start:

- Phase 1: Data Familiarization - pending
- Phase 2: Pattern Discovery - pending
- Phase 3: Anomaly Investigation - pending
- Phase 4: Insight Generation - pending
- Phase 5: Question Formulation - pending

Update status as you progress. Mark phases complete ONLY after checkpoint verification.

---

Phase 1: Data Familiarization

CHECKPOINT: Before proceeding, you MUST have:

• Catalogued all tables and their apparent purposes
• Documented schema (columns, types, relationships)
• Assessed data quality (completeness, ranges, distributions)
• Identified temporal coverage and granularity
• Saved to 01 - data-familiarization.md

Instructions

1. Understand what data you have

Create analysis/[session-name]/01 - data-familiarization.md with: ./templates/01-data-familiarization.md

2. Use the understanding-data component skill

   • Profile the database systematically
   • Don't skip quality checks - surprises are common
   • Look for obvious data issues that would affect exploration

3. Resist premature pattern-hunting

   • Don't start with "I wonder if weekends are different"
   • First understand WHAT data you have, THEN explore patterns
   • If you catch yourself forming hypotheses, note them for Phase 2 but finish familiarization first

Common Rationalization: "I can see the tables, I don't need detailed familiarization" Reality: Skipping familiarization leads to missing important context about data quality, coverage, and structure.

Common Rationalization: "I'll explore patterns while I familiarize myself with the data" Reality: Mixing familiarization and pattern-hunting creates confusion. Separate concerns.

---

Phase 2: Pattern Discovery

CHECKPOINT: Before proceeding, you MUST have:

• Explored temporal patterns (trends, seasonality, cycles)
• Explored segmentation patterns (groups, categories, clusters)
• Explored relationship patterns (correlations, associations)
• Documented each exploration with rationale, query, results, observations
• Created separate files for each exploration vector
• Files saved as 02-temporal-patterns.md, 03-segmentation-patterns.md, 04-relationship-patterns.md

Instructions

1. Explore SYSTEMATICALLY along three vectors

You MUST explore all three vectors, even if some seem less promising. Surprises often come from unexpected places.

Vector 1: Temporal Patterns

Create analysis/[session-name]/02-temporal-patterns.md with: ./templates/02-temporal-patterns.md

Vector 2: Segmentation Patterns

Create analysis/[session-name]/03-segmentation-patterns.md with: ./templates/03-segmentation-patterns.md

Vector 3: Relationship Patterns

Create analysis/[session-name]/04-relationship-patterns.md with: ./templates/04-relationship-patterns.md

2. Be systematic, not random

   • Explore ALL three vectors (temporal, segmentation, relationship)
   • Don't skip a vector just because first query seems uninteresting
   • Patterns often hide in places you don't expect

3. Separate observation from interpretation

   • In each analysis, state FACTS (what the numbers show)
   • Save interpretation for Phase 4 (Insight Generation)
   • Resist the urge to explain patterns yet - just catalog them

4. Use visualizations liberally

   • Text-based tables, ASCII charts, markdown formatting
   • Help patterns become visible
   • Refer to creating-visualizations component skill

Common Rationalization: "I found an interesting pattern, I'll just focus on that" Reality: Focusing on first interesting pattern causes you to miss better patterns elsewhere. Complete all three vectors.

Common Rationalization: "This dimension looks boring, I'll skip it" Reality: "Boring" dimensions often hide surprising patterns. Explore systematically.

Common Rationalization: "I'll combine all patterns into one analysis file" Reality: Separate files by vector creates structure and makes findings easy to locate.

---

Phase 3: Anomaly Investigation

CHECKPOINT: Before proceeding, you MUST have:

• Identified 3-5 specific anomalies, outliers, or unexpected patterns from Phase 2
• Investigated each anomaly with follow-up queries
• Determined if anomalies are data quality issues or real phenomena
• Documented findings
• Saved to 05 - anomaly-investigation.md

Instructions

1. Review Phase 2 for anomalies

Go back through temporal, segmentation, and relationship explorations and identify:

• Unexpected spikes or drops
• Outlier values or segments
• Counterintuitive patterns
• Violations of expected business logic
• Inconsistencies or irregularities

2. Investigate each anomaly

Create analysis/[session-name]/05 - anomaly-investigation.md with: ./templates/05-anomaly-investigation.md

3. Distinguish data quality from real patterns

   • NULL values, duplicates, data entry errors → data quality issues
   • Seasonal events, one-time promotions, external shocks → real phenomena
   • When unclear, state that explicitly

4. Don't ignore anomalies

   • Anomalies are often where the most interesting insights hide
   • Investigate thoroughly, don't dismiss as "probably nothing"
   • Data quality issues found now save headaches later

Common Rationalization: "That spike is probably a holiday, I'll ignore it" Reality: VERIFY your assumption. "Probably" isn't good enough. Check.

Common Rationalization: "This anomaly is a data quality issue, I'll just exclude it" Reality: Document what you're excluding and why. Future analysts need to know.

Common Rationalization: "I found 20 anomalies, I'll investigate them all" Reality: Focus on the 3-5 most significant. You're exploring, not auditing every data point.

---

Phase 4: Insight Generation

CHECKPOINT: Before proceeding, you MUST have:

• Reviewed all patterns and anomalies from Phases 2-3
• Identified 3-5 insights that are actionable, surprising, or meaningful
• For each insight, documented: finding, significance, context, caveats
• Assessed confidence level for each insight
• Saved to 06 - insights.md

Instructions

1. Extract insights from patterns

Not every pattern is an insight. An insight must be:

• Actionable: Suggests a decision or further investigation
• Surprising: Non-obvious, counter to expectations, or revealing hidden structure
• Meaningful: Materially significant magnitude or impact

2. Document insights systematically

Create analysis/[session-name]/06 - insights.md with: ./templates/06-insights.md

3. Be selective

   • Don't try to turn every pattern into an insight
   • 3-5 high-quality insights beat 20 marginal observations
   • If you have more than 5, rank them and focus on top 5

4. Quantify magnitude

   • "Sales vary by region" is not enough
   • "Top region has 3x sales of bottom region" is specific
   • Numbers make insights actionable

5. Assess confidence honestly

   • High confidence: Strong evidence, large sample, consistent pattern, clear interpretation
   • Medium confidence: Clear pattern but causation unclear OR limited time window OR potential confounds
   • Low confidence: Interesting but small sample OR inconsistent OR multiple plausible explanations

Common Rationalization: "Every pattern I found is an insight" Reality: Most patterns are noise or expected. Be selective. Insights must clear the bar: actionable, surprising, meaningful.

Common Rationalization: "I'll just state the pattern and let the user figure out if it matters" Reality: Your job is to interpret. Explain WHY the pattern matters and WHAT it suggests.

Common Rationalization: "I'm very confident in this insight" Reality: Exploratory analysis rarely produces high confidence. Be honest about uncertainty and limitations.

---

Phase 5: Question Formulation

CHECKPOINT: Before proceeding, you MUST have:

• Generated 3-5 specific, answerable questions based on insights
• For each question, identified: what process to use, what data is needed, why it matters
• Prioritized questions by potential value
• Saved to 07 - next-questions.md
• Updated 00 - overview.md with exploration summary

Instructions

1. Convert insights into questions

Each insight should suggest 1-2 follow-up questions. Good questions:

• Build on what you learned in exploration
• Are specific enough to be answerable
• Have clear business value
• Can be addressed with available (or obtainable) data

2. Document questions for follow-up

Create analysis/[session-name]/07 - next-questions.md with: ./templates/07-next-questions.md

3. Update exploration overview

Update analysis/[session-name]/00 - overview.md by adding content from: ./templates/overview-summary.md

4. Final verification checklist
   • Explored all three vectors (temporal, segmentation, relationship)
   • Investigated significant anomalies
   • Generated 3-5 qualified insights (actionable, surprising, meaningful)
   • Formulated specific follow-up questions
   • Prioritized questions by value
   • Updated overview with summary
   • Communicated findings to user

Common Rationalization: "I found interesting patterns, I'm done" Reality: Exploration isn't complete until you've formulated what to investigate next. Always end with questions.

Common Rationalization: "I'll just suggest broad areas to explore further" Reality: Be specific. "Investigate customer behavior" is not actionable. "Compare weekend vs weekday sales per operating hour using comparative-analysis skill" is actionable.

Common Rationalization: "I'll list every possible question I can think of" Reality: Focus on the 3-5 highest-value questions. Too many options create decision paralysis.

---

Common Rationalizations

"I can see interesting patterns already, I'll skip data familiarization"

Why this is wrong: Without understanding data quality, coverage, and structure, your pattern discoveries may be artifacts or noise.

Do instead: Complete Phase 1 fully. Familiarization prevents false discoveries and wasted effort.

"This exploration vector looks boring, I'll skip it"

Why this is wrong: The most surprising insights often come from places you didn't expect. "Boring" dimensions frequently hide interesting patterns.

Do instead: Explore ALL three vectors (temporal, segmentation, relationship) systematically. Be comprehensive.

"I found one great insight, that's enough"

Why this is wrong: One insight doesn't exhaust a dataset. You likely missed other valuable patterns.

Do instead: Continue systematic exploration. Aim for 3-5 insights across different dimensions.

"Every pattern I found is an insight"

Why this is wrong: Most patterns are noise, expected, or immaterial. Calling everything an insight dilutes the valuable discoveries.

Do instead: Apply strict criteria: actionable, surprising, meaningful. Be selective.

"I'll just describe patterns and let the user decide if they're important"

Why this is wrong: Your job is interpretation, not just data reporting. Users expect you to identify what matters and why.

Do instead: Assess significance, provide context, explain business implications. Do the analytical thinking.

"I'm very confident in these exploratory findings"

Why this is wrong: Exploratory analysis generates hypotheses, not confirmations. Patterns need validation through targeted investigation.

Do instead: Be honest about confidence levels. Exploratory findings are typically medium-low confidence until validated.

"I found interesting patterns, analysis is complete"

Why this is wrong: Exploration is the beginning, not the end. The goal is to identify what to investigate deeply.

Do instead: Always complete Phase 5. Convert insights into specific, answerable questions for follow-up.

"I'll combine all patterns into one big report"

Why this is wrong: Mixing temporal, segmentation, and relationship analyses creates confusion and makes findings hard to locate.

Do instead: Separate files by exploration vector (02-temporal, 03-segmentation, 04-relationship). Clear structure aids comprehension.

"This anomaly is probably a data error, I'll just exclude it"

Why this is wrong: Assumptions about anomalies are often wrong. "Probably" isn't good enough. Also, excluding data without documentation creates reproducibility issues.

Do instead: Investigate anomalies in Phase 3. Document what you exclude and why. Verify your assumptions.

"I have 15 follow-up questions, I'll list them all"

Why this is wrong: Too many options create decision paralysis. Not all questions are equally valuable.

Do instead: Prioritize ruthlessly. Focus on 3-5 highest-value questions. Help the user know where to start.

---

Summary

This skill ensures systematic, thorough exploration of unfamiliar datasets by:

1. Familiarizing with data first: Understand structure, quality, and coverage before pattern-hunting
2. Exploring systematically: Three vectors (temporal, segmentation, relationship) ensure comprehensive discovery
3. Investigating anomalies: Surprises and outliers often contain the most valuable insights
4. Generating selective insights: Apply strict criteria (actionable, surprising, meaningful) to separate signal from noise
5. Formulating specific questions: Convert insights into answerable questions for deeper investigation
6. Prioritizing next steps: Help users know what to investigate first and why

Follow this process and you'll discover what's truly interesting in unfamiliar data, avoid random pattern-chasing, and identify high-value questions for targeted investigation.