Question Framing

Skill: Question Framing

Purpose

Structure analytical questions using the Question Ladder framework so every analysis starts with a clear decision context, measurable success criteria, and testable hypotheses.

When to Use

Apply this skill when starting any new analysis, when a user asks a vague question ("How are we doing?"), or when an analysis request lacks decision context. Always frame before analyzing.

Instructions

Pre-flight: Load Learnings

Before executing, check .knowledge/learnings/index.md for relevant entries:

• Read the file. If it doesn't exist or is empty, skip silently.
• Scan for entries under "Question Framing" and "General" headings (or related categories like "Business Context", "Methodology Notes").
• If entries exist, incorporate them as constraints or context for this execution.
• Never block execution if learnings are unavailable.

The Question Ladder

Every analytical question climbs four rungs:

GOAL        → What business outcome are we trying to achieve?
DECISION    → What specific decision will this analysis inform?
METRIC      → What will we measure to inform that decision?
HYPOTHESIS  → What do we expect to find, and why?

The rule: Never start analyzing data until you can state all four rungs. If the requester only gives you a goal ("improve retention"), your first job is to climb the ladder before touching data.

Framing Process

Step 1: Extract the decision Ask: "What will you DO differently based on the answer?"

• If the answer is "nothing" or "I'm just curious" → this is reporting, not analysis. Offer two paths:
  • Path A: Quick stat/dashboard (if truly no decision)
  • Path B: Clarify decision context first, then frame properly
• If the answer is a specific action → you have a decision. Proceed to Step 2.

Step 2: Define success criteria Ask: "How will you know the analysis answered your question?"

• The answer should be specific: "If conversion rate dropped >10% in segment X, we'll prioritize a fix"
• Not vague: "We'll understand our users better"
• Success criteria should include specific thresholds, conditions, or decision rules

Step 3: Form testable hypotheses Ask: "What do you think is happening, and why?"

• Good: "I think mobile conversion dropped because the checkout redesign broke on small screens"
• Bad: "I think things are bad"
• Extract hypotheses from vague statements - e.g., "cart abandonment is high" → "abandonment >20% due to checkout friction at payment stage"

Step 4: Identify data requirements Ask: "What data do we need, and do we have it?"

• Map each hypothesis to specific metrics, segments, and time ranges
• Check if the data exists by reviewing schema documentation (don't query the database - just check availability)
• Flag gaps early: "We need funnel step events, but only have order status"
• Document what CAN and CANNOT be answered with available data

Step 5: Produce the Question Brief Write the brief using the template below. Save it to question_brief.md in the working folder, or present it inline.

• DO NOT proceed to analysis after writing the brief
• DO NOT run SQL queries or call analysis agents
• Hand off to the next phase (exploration/analysis) after the brief is approved

Good vs. Bad Questions

Impact × Feasibility Prioritization

When multiple questions emerge, prioritize:

                    HIGH IMPACT
                        │
          ┌─────────────┼─────────────┐
          │   DO FIRST   │   PLAN FOR  │
          │  (Quick win)  │  (Strategic) │
HIGH      │               │              │
FEASIBILITY ──────────────┼──────────────── LOW
          │               │              │ FEASIBILITY
          │   DO IF TIME  │    SKIP     │
          │  (Nice to have)│  (Not worth) │
          └─────────────┼─────────────┘
                        │
                    LOW IMPACT

Impact criteria:

• Revenue/cost implication >$100K → High
• Affects >10% of users → High
• Informs a decision being made this quarter → High
• Curiosity-driven, no pending decision → Low

Feasibility criteria:

• Data exists and is clean → High
• Can be answered in <4 hours → High
• Requires new instrumentation → Low
• Requires data from another team → Low

Output Format: Question Brief (Ladder + Analysis Design Spec)

The Question Brief is the concrete artifact the Ladder produces. Climb the Ladder first, then fill the seven spec fields. Every field is required; if you cannot fill one, ask the user. Save it to question_brief.md in the working folder (or present it inline for quick asks).

# Question Brief: [Title]
## Date: [YYYY-MM-DD]

### Business Context
[2-3 sentences: what's happening, why this matters now]

### The Question Ladder
| Rung | Statement |
|------|-----------|
| **Goal** | [Business outcome] |
| **Decision** | [Specific action this informs] |
| **Metric** | [What we'll measure] |
| **Hypothesis** | [What we expect to find and why] |

### 1. Question
What are we trying to answer?
[A specific, testable question, sharpened from the Ladder]

### 2. Decision
What will this analysis inform?
[A concrete action the team will take based on the answer]
[If the answer is "nothing specific", this may be reporting, not analysis. Confirm with the user.]

### 3. Data Needed
| Data | Source | Available? | Notes |
|------|--------|-----------|-------|
| [metric/field] | [table/system] | Yes/No/Partial | [gaps, quality concerns] |

### 4. Dimensions
What should we segment or decompose by?
- [Dimension 1]: [why: what would different values tell us?]
- [Dimension 2]: [why]
- [Dimension 3]: [why]

### 5. Time Range & Granularity
- **Period:** [start date to end date]
- **Granularity:** [daily / weekly / monthly]
- **Comparison:** [vs. prior period / vs. same period last year / vs. benchmark]

### 6. Output Format
What deliverable does the user need?
- [ ] Quick answer (1-2 sentences + supporting number)
- [ ] Analysis report (structured findings with charts)
- [ ] Presentation deck (slides for stakeholders)
- [ ] Data table (for further analysis by the user)

### 7. Success Criteria
How will we know the analysis answered the question?
[Specific, falsifiable conditions, e.g. "Identify which segment drove >50% of the decline"]

### Priority
- **Impact:** [High/Medium/Low, with justification]
- **Feasibility:** [High/Medium/Low, with justification]
- **Recommendation:** [Do First / Plan For / Do If Time / Skip]