Decide Actions

Decide 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 configure-tasks skill)
│
└─ Just informing/acknowledging?
    └─ → RESPOND WITHOUT DISCOVERY

Detailed Decision Matrix

Disambiguation — Segmentation vs Funnel:

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

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:

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:

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

Respond Without Creating a 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 configure-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:

Discovery Freshness Rules

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

Analysis Approach Selection

Question Type → Approach

Complexity Assessment

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:

Reference Files

• Insight type selection
• Duplicate prevention
• Discovery workflow