deciding-actions

Deciding Actions

Quick Start

This skill provides decision frameworks for:

1. Choosing the right insight type (funnel, retention, semantic, segmentation, SQL)
2. Choosing the right task type (anomaly_detection, forecast, monitor)
3. Deciding discovery actions (new/update/skip)
4. Avoiding duplicate discoveries
5. Selecting analysis approaches

When to Use This Skill

• Choosing between funnel, retention, semantic, segmentation, or SQL insights
• Choosing a task type for automated monitoring (anomaly_detection, forecast, monitor)
• Deciding whether to create a new discovery
• Checking for duplicate discoveries
• Selecting the right analysis method
• Planning discovery workflow

Insight Type Decision Matrix

How to Use

1. Match the user's question against the decision tree below
2. If ambiguous, check the signal matrix for matching phrases
3. If still ambiguous, use the disambiguation blocks to resolve the overlap
4. Cross-check against the common misclassifications before creating

Quick Decision Tree

User Question
│
├─ About conversion/steps/flow?
│   └─ → FUNNEL INSIGHT
│
├─ About whether users come back after an event?
│   └─ → RETENTION INSIGHT
│
├─ About metrics/dimensions/trends?
│   └─ → SEMANTIC INSIGHT
│
├─ Complex/custom/joins needed?
│   └─ → SQL INSIGHT
│
├─ About comparing event metrics across segments/cohorts?
│   └─ → SEGMENTATION INSIGHT
│
├─ Need automated recurring analysis?
│   └─ → TASK (see configuring-tasks skill)
│
└─ Just informing/acknowledging?
    └─ → FYI DISCOVERY

Detailed Decision Matrix

Signal	Funnel	Retention	Semantic	SQL	Segmentation	FYI
"conversion rate"	✓✓✓
"drop-off"	✓✓✓
"steps to purchase"	✓✓✓
"user journey"	✓✓✓
"stuck at step/level"	✓✓✓
"progression from X to Y"	✓✓✓
"did X but not Y"	✓✓✓
"come back"		✓✓✓
"return after"		✓✓✓
"retained"		✓✓✓
"churn"		✓✓✓
"how many"			✓✓✓
"trend over time"			✓✓✓
"breakdown by"			✓✓✓
"compare periods"			✓✓✓
"join tables"				✓✓✓
"custom calculation"				✓✓✓
"raw data"				✓✓✓
"complex query"				✓✓✓
"users who [have property]"					✓✓✓
"cohort of"					✓✓✓
"segment where"					✓✓✓
"acknowledge"						✓✓✓
"got it"						✓✓✓
"thanks"						✓✓✓

Disambiguation — Segmentation vs Funnel:

The phrase "users who" is ambiguous. Apply this test:

Pattern	Type	Why
"users who have property X"	Segmentation	Defining a cohort for behavioral comparison
"users who did event A then event B"	Funnel	Sequential event analysis
"event count by plan/device/source"	Segmentation	Event metric comparison across property values
"users stuck at step/level X"	Funnel	Step-to-step progression
"users who completed X but not Y"	Funnel	Measuring drop-off between steps
"users in segment/group X"	Segmentation	Pre-defined cohort

Key test: Is the finding about comparing event behavior across cohorts/properties (→ segmentation) or movement through ordered steps (→ funnel)?

Disambiguation — Semantic vs SQL:

Both produce metric values. Apply this test:

Factor	Semantic	SQL
Metric/dimension exists in semantic model	✓
Requires joins across tables		✓
Custom calculation or formula		✓
Data not modeled in semantic layer		✓
Standard breakdown (e.g., revenue by region)	✓

Key test: Does the semantic model already expose this metric and dimension? Yes → Semantic. No → SQL. When unsure, check the semantic model first.

Disambiguation — Funnel vs Retention:

Both involve user events over time. Apply this test:

Pattern	Type	Why
"users who go from A to B"	Funnel	Sequential step progression
"users who come back after A"	Retention	Return behavior over time
"drop-off between steps"	Funnel	Measuring where users stop in a sequence
"churn after event X"	Retention	Measuring who doesn't return

Key test: Is the finding about moving through a sequence of steps (→ funnel) or coming back after a starting event (→ retention)?

When to Use Each Type

Use FUNNEL INSIGHT When

• User asks about conversion rates
• Question involves sequential steps or progression
• Analyzing user journey/flow
• Finding where users drop off or get stuck
• Measuring completion rates between stages
• Comparing progression across levels, tiers, or milestones

Keywords: conversion, funnel, steps, journey, drop-off, flow, complete, abandon, stuck, progression, level, stage, bottleneck

Use RETENTION INSIGHT When

• Analyzing whether users return after a starting event
• Measuring churn or repeat behavior over time
• Comparing retention across cohorts or time periods
• Tracking if users who did event A come back to do event B

Keywords: retention, churn, come back, return, repeat, re-engage, day 1/7/30

Use SEMANTIC INSIGHT When

• User asks about metrics/KPIs
• Question involves trends over time
• Needs breakdown by dimension
• Standard analytics questions
• Comparing time periods

Keywords: how many, trend, breakdown, compare, metric, daily, weekly, growth

Use SQL INSIGHT When

• Semantic model doesn't have needed data
• Complex joins required
• Custom calculations needed
• Raw data exploration
• Ad-hoc analysis

Keywords: join, custom, raw, specific table, complex, calculate

Use SEGMENTATION INSIGHT When

• Comparing event metrics across cohorts (e.g., feature usage by plan, device, or region)
• Breaking down events by event, user, or session properties
• Segmenting behavior over time without requiring ordered steps
• Building cohorts for targeting, comparison, or further analysis
• Filtering users by dimensions like device, plan, region, etc.

Keywords: segment, cohort, group of, target, users with, users in

Use FYI DISCOVERY When

• Acknowledging user input
• No analysis needed
• Informational response
• Status update
• Simple confirmation

Keywords: thanks, got it, understood, noted, will do

Task Type Decision Matrix

When the user needs automated, recurring analysis rather than a one-off insight, choose a task type:

User wants automation
│
├─ Detect outliers/anomalies in an Insight?
│   └─ → anomaly_detection task
│
├─ Project future values from an Insight?
│   └─ → forecast task
│
└─ Open-ended AI analysis of an Insight/Dashboard?
    └─ → monitor task

See the configuring-tasks skill for full task creation workflow.

Discovery Action Decision Matrix

Create vs Update vs Skip

Is this finding new?
│
├─ YES: Does similar discovery exist?
│   │
│   ├─ NO → CREATE NEW
│   │
│   └─ YES: Is new info significantly different?
│       │
│       ├─ YES → CREATE NEW (with reference to previous)
│       │
│       └─ NO → SKIP (already covered)
│
└─ NO: Is it a follow-up to previous?
    │
    ├─ YES → CREATE NEW (as follow-up)
    │
    └─ NO → SKIP (redundant)

Duplicate Detection Checklist

Before creating a discovery, check:

Check	Action if True
Same metric, same time range, same finding?	SKIP
Same insight within last 24 hours?	SKIP
Same topic, minor variation?	SKIP
Contradicts recent discovery?	CREATE (with explanation)
Adds significant new context?	CREATE
User explicitly asked again?	CREATE

Discovery Freshness Rules

Discovery Age	Same Topic Action
< 1 hour	Always skip unless contradicts
1-24 hours	Skip unless significant new info
1-7 days	Create if adds value
> 7 days	Treat as fresh topic

Semantic Model Check

Before SQL, Check Semantic

Need data?
│
├─ Check semantic model first
│   │
│   ├─ Dimension/measure exists? → Use SEMANTIC
│   │
│   └─ Not available?
│       │
│       ├─ Can be added to model? → Consider adding, then SEMANTIC
│       │
│       └─ One-off need? → Use SQL

Semantic vs SQL Decision

Factor	Prefer Semantic	Prefer SQL
Reusability	✓
Consistency	✓
Performance	✓
Flexibility		✓
Complex joins		✓
One-off analysis		✓
Standard metrics	✓
Custom calculations		✓

Analysis Approach Selection

Question Type → Approach

Question Pattern	Primary Approach	Fallback
"Why did X happen?"	Semantic breakdown	SQL drill-down
"How is X performing?"	Semantic trend	Dashboard
"Who are the users that..."	Segmentation	SQL filter
"What's the conversion..."	Funnel	SQL with steps
"Do users come back after..."	Retention	Funnel fallback
"Compare A vs B"	Semantic comparison	SQL union
"Predict/forecast"	Forecast task	Semantic trend as fallback

Complexity Assessment

Complexity	Approach	Notes
Simple metric	Semantic	Direct query
Metric + filter	Semantic	Add dimension filter
Metric + breakdown	Semantic	Group by dimension
Multi-step analysis	Funnel or SQL	Depends on data
Cross-table	SQL	Joins required
Historical comparison	Semantic	Time dimension

Avoiding Common Mistakes

Don't Create Discovery When

• Already said the same thing recently
• Information is trivial/obvious
• User didn't ask for insight
• Just acknowledging without adding value
• Repeating what user already knows

Do Create Discovery When

• New actionable insight found
• Significant change detected
• User explicitly requested analysis
• Important pattern identified
• Anomaly requires attention

Quality Gates

Before creating any discovery:

1. Novelty: Is this new information?
2. Value: Does this help the user?
3. Accuracy: Is the data correct?
4. Actionable: Can user do something with this?
5. Timing: Is now the right time?

Common Misclassifications

Findings that are frequently assigned the wrong insight type:

Finding	Wrong Choice	Right Choice	Why
"Users stuck at step/level X"	Segmentation	Funnel	Step progression = sequential analysis
"Drop-off between A and B"	SQL	Funnel	Sequential steps with conversion
"Users who did X but not Y"	Segmentation	Funnel	Sequential dependency between events
"Metric broken down by property"	SQL	Semantic	Standard breakdown = use semantic model
"Metric X by dimension Y"	SQL	Semantic	Dimension likely exists in model
"Users with property X"	Funnel	Segmentation	Attribute-based group, not a flow
"Do users come back after X?"	Funnel	Retention	Return behavior, not step progression
"Churn after event X"	Segmentation	Retention	Measuring who doesn't return over time

Reference Files

• Insight type selection
• Duplicate prevention
• Discovery workflow