references/synthesize-research-report/phases/phase4-synthesis.md

Phase 4: Synthesis & Interpretation

You are a sub-agent executing Phase 4 of a qualitative research synthesis. This is the most intellectually demanding phase — you must integrate everything into prioritized findings, data-driven personas, outcome-oriented opportunities, and a curated evidence bank. This phase transforms analysis into actionable insight.

Context

You will receive:

analysis/config.md — configuration including research goal, persona type, prioritization framework
analysis/phase1-familiarization/consolidated-observations.md — initial observations
analysis/phase1-familiarization/batch-{n}-memos.md — all interview memos
analysis/phase2-coding/codebook.md — complete codebook
analysis/phase2-coding/coded-excerpts/ — all coded excerpts
analysis/phase3-themes/themes.md — theme definitions
analysis/phase3-themes/frequency-matrix.md — prevalence data
analysis/phase3-themes/pattern-analysis.md — cross-cutting patterns
analysis/phase3-themes/co-occurrence.md — code/theme clusters
analysis/phase3-themes/phase3-summary.md — theme summary and recommendations

Your Outputs

Write all outputs to analysis/phase4-synthesis/.

Task 1: Construct Personas

Step 1: Confirm Persona Type (`personas/persona-type-rationale.md`)

Review config.md for the pre-selected persona type. Validate it against what the data actually shows:

If data shows...	Best persona type
Distinct usage patterns and skill levels	Behavioral
Distinct mindsets, motivations, or values	Attitudinal
Distinct goals regardless of demographics	Goal-Based
Need to communicate to non-researchers	Narrative
Multiple organizational roles in the buying/using chain	Ecosystem
Limited data, need team alignment fast	Proto-Personas
Designing a bot/AI personality	System/VUI

If the data suggests a different type than what was configured, document why and switch.

Step 2: Cluster Participants (`personas/clustering-analysis.md`)

Cluster by reasoning and behavior, not demographics. Group participants by their patterns of action, domain knowledge, and motivations (e.g., "necessity-oriented" vs. "entertainment-oriented") rather than age or gender. Demographics are only relevant when they directly change how a person interacts with the product.

Identify the clustering dimensions based on persona type:

Behavioral: Cluster by usage frequency, feature adoption, skill level, workflow patterns Attitudinal: Cluster by values, motivations, risk tolerance, decision-making style Goal-Based: Cluster by primary objectives, success criteria, jobs-to-be-done Narrative: Cluster by life/work context, journey stage, relationship to product/domain Ecosystem: Cluster by organizational role, decision authority, success metrics

# Clustering Analysis

## Clustering Dimensions
- [Dimension 1]: [Definition + spectrum]
- [Dimension 2]: [Definition + spectrum]

## Participant Mapping

| Participant | Dim 1 | Dim 2 | Cluster |
|-------------|-------|-------|---------|
| P01 | [position] | [position] | A |
| P02 | [position] | [position] | B |

## Identified Clusters
### Cluster A: [Working name]
- Participants: [list]
- Shared characteristics: [what unites them]
- Size: [count] ([%] of sample)

### Cluster B: [Working name]
...

## Rationale for Number of Personas
[Why this number? What would be lost by merging? What would be gained by splitting?]
Auto-determine the count: typically 3-5. Fewer than 3 suggests under-differentiation. More than 6 suggests overlap.

Step 3: Write Individual Personas (`personas/persona-{n}-{name}.md`)

The exact template depends on persona type. Core structure applies three principles:

"Differences that matter" filter: Only include details that change how this person interacts with the product or domain. If a trait (age, location, job title) is irrelevant to their behavior, omit it. Every detail earns its place.
Narrative glue: Even for non-Narrative persona types, compose a brief composite that blends attributes from several similar participants into a cohesive character. The persona should feel like a real person, not a data table.
Required anchors: Every persona MUST have a defining quote (one quote that personifies their mindset) and 3-5 high-level goals (what they care about at the end of the day).

# Persona {N}: [Name]

**Type**: [Behavioral / Attitudinal / Goal-Based / Narrative / Ecosystem]

**Defining Quote**:
> "[The single quote that best personifies this persona's mindset]"

## Who They Are
- [Primary identifying characteristics relevant to the persona type — only "differences that matter"]
- [Context: role, environment, experience level, relationship to product/domain]
- [How they arrived at their current situation]

## High-Level Goals
1. [What they care about at the end of the day — outcome-level, not feature-level]
2. [Goal 2]
3. [Goal 3]
(3-5 goals)

## [Section varies by type]:

### For Behavioral Personas:
**How They Work/Use**:
- Frequency: [how often]
- Primary workflows: [what they do]
- Tools and integrations: [what else they use]
- Skill level: [novice → expert]
- Workarounds: [what they've hacked together]

### For Attitudinal Personas:
**What They Believe**:
- Core values: [what drives decisions]
- Attitude toward [domain]: [their stance]
- Risk tolerance: [how they handle uncertainty]
- Decision-making style: [how they evaluate options]

### For Goal-Based Personas:
**What They're Trying to Accomplish**:
- Primary job-to-be-done: [the outcome they seek]
- Success criteria: [how they know they've succeeded]
- Current approach: [how they try to achieve this today]
- Barriers: [what blocks them]

### For Narrative Personas:
**A Day in Their Life**:
[2-3 paragraph narrative showing a typical scenario — grounded in real interview data, composited across cluster members. Show, don't tell.]

### For Ecosystem Personas:
**Their Role in the System**:
- Decision authority: [what they can approve/block]
- Success metrics: [what they're measured on]
- Information needs: [what they need to know]
- Relationship to other personas: [how they interact]

## Key Needs & Pain Points
1. [Need/Pain 1]: [Evidence — theme reference + brief quote]
2. [Need/Pain 2]: [Evidence]
3. [Need/Pain 3]: [Evidence]

## Representative Quotes (2-4)
> "[Quote that captures their voice]"
> — Context: [situation]

## How They Relate to Key Themes
- [Theme X]: [How this persona experiences this theme]
- [Theme Y]: [How this persona experiences this theme differently than other personas]

## Participants Represented
[List of participant IDs in this cluster — for traceability]

Task 2: Synthesize Prioritized Findings (`findings.md`)

Selecting the Prioritization Framework

Read config.md for the research goal and apply the corresponding framework:

Criticality Scoring (for tactical usability fixes):

Score each finding: Severity (1-4) x Frequency (% of sample)
Rank by composite score

Opportunity Scoring (for market gaps & innovation):

Score each finding: Importance (how much users care) vs. Satisfaction (how well current solutions work)
Prioritize where importance is high but satisfaction is low

Impact/Effort Matrix (for MVP/sprint scoping):

Score each finding: Customer Impact (high/medium/low) vs. Implementation Effort (high/medium/low)
Prioritize "quick wins" (high impact, low effort)

Hypothesis Canvas (for high-risk discovery):

Score each finding: Risk if wrong vs. Perceived Value
Prioritize "leap of faith" assumptions

Opportunity Solution Tree (for product strategy):

Map findings as opportunities under desired outcomes
Prioritize branches with strongest evidence and clearest path to solutions

Document your framework application in prioritization-rationale.md.

Important: Treat prioritization as a "two-way door" — a reversible decision. If new data or stakeholder feedback shows a high-priority finding is less impactful than expected, the team must be able to course-correct without delay. Note this explicitly in the rationale.

Finding Tiers

After scoring, sort findings into three tiers:

Tier	Label	Criteria	Report Treatment
1	Must Know	High severity/impact, high frequency, high confidence	Detailed treatment, leads the report, demands immediate action
2	Should Know	Moderate impact or frequency, solid evidence	Standard treatment in the body of the report
3	Nice to Know	Lower frequency, exploratory, or lower confidence	Brief treatment, may move to appendix or open questions

Finding Structure

Aim for 5-8 key findings. Each finding must appear in at least 2 interviews.

## Finding [N]: [One clear sentence stating the insight]

**Tier**: [Must Know / Should Know / Nice to Know]

**Priority**: [Rank from prioritization framework + score if applicable]

**Prevalence**: [X] of [Y] participants ([Z]%)

**Theme(s)**: [Which themes from Phase 3 this finding synthesizes]

**Persona impact**: [Which personas are most affected and how]

**Evidence**:

*Luminous exemplar* (the single most powerful quote):
> "[Verbatim quote]"
> — [Participant ID], [brief context]
> Why this quote: [What makes it analytically illuminating]

*Anchor candidate* (for extended treatment):
- [Participant ID]: [1-sentence reason they could illustrate this finding in depth]

*Supporting quotes (echoes)* showing prevalence:
> "[Short quote]" — [Participant ID]
> "[Short quote]" — [Participant ID]
> "[Short quote]" — [Participant ID]

**Variation**: [How does this finding manifest differently across participants or personas?]

**Negative cases**: [Participants who DON'T show this pattern — why not?]

**Confidence level**: [High / Medium / Low] — [Basis: number of sources, behavioral vs stated, triangulation]

Task 3: Build the Evidence Bank (`evidence-bank.md`)

The evidence bank is NOT a dump of every quote — it is a curated "case file" containing only the meaningful bits of data that comprise your findings. A single well-chosen quote can be "worth ten thousand words" in driving team empathy.

Curation Principles

Stickiness: Select quotes that will resonate in a meeting room. The best quotes make the listener feel the participant's experience — they create empathy, not just understanding.
Traceability: Every quote must be labeled with its original location (participant ID, transcript page/line number or timestamp) so the full context can be re-examined if a conclusion is challenged.
Paralinguistic cues: Where available from transcripts, annotate quotes with non-verbal signals — tone, pauses, laughter, sighing, emphasis. These cues carry meaning that words alone miss.

# Evidence Bank

## Finding [N]: [Name]

### Top Quotes (ranked by analytical power)

1. **Luminous exemplar**:
   > "[Full quote]"
   > — [Participant ID] ([demographics/context]) | Source: [transcript page/line or timestamp]
   > Paralinguistic cues: [tone, pauses, emphasis — if available]
   > Analytical value: [Why this quote does work — what it shows that explanation alone cannot]
   > Stickiness: [Why this quote will resonate with stakeholders]

2. **Anchor material** (extended excerpt for deep illustration):
   > "[3-6 sentence excerpt showing reasoning, emotion, or sequence]"
   > — [Participant ID] ([demographics/context]) | Source: [transcript location]
   > Paralinguistic cues: [if available]

3-5. **Echo quotes** (showing breadth):
   > "[1-2 sentence quote]" — [Participant ID] | Source: [transcript location]

### Counter-Evidence
> "[Quote from negative case]" — [Participant ID] | Source: [transcript location]
> How this was addressed: [explanation]

Task 4: Map Opportunities (`opportunities.md`)

Transform findings into outcome-oriented opportunity statements. Frame as customer needs, pain points, or desires — never as feature requests or solutions.

Opportunity Formulation

Use one of two formats depending on precision needed:

Standard format: "Enable [persona/users] to [desired outcome] without [current barrier/pain]"

Precision format (Direction + Measure + Object + Clarifier): "[Minimize/Maximize/Increase/Reduce] the [time/effort/cost/risk] to [action/object] [for context/clarifier]"

Example: "Minimize the time it takes to gather documents for sharing with colleagues"

The precision format is better when the opportunity needs to be measurable or directly translatable into requirements.

Hierarchical Structure

Opportunities should be structured as a tree, not a flat list. Large, project-sized opportunities must be broken into smaller, more solvable sub-opportunities (leaf-nodes). When prioritizing, always prefer to address a leaf-node (an opportunity with no children) to ensure the team delivers iterative value quickly.

# Opportunity Areas

## Opportunity [N]: [Outcome-oriented statement — standard or precision format]

**Level**: [Root / Branch / Leaf]

**Parent opportunity**: [Reference to parent, if this is a branch or leaf — "None" if root]

**Derived from**: Finding [X], Finding [Y]

**Affected personas**: [Which personas + how specifically]

**Current state**: [How users handle this today — workarounds, pain, avoidance]

**Desired state**: [What success looks like from the user's perspective]

**Evidence strength**: [High/Medium/Low] — [Basis]

**Priority**: [From the prioritization framework applied to findings]

**Sub-opportunities** (if this is not a leaf):
- [N.1]: [Sub-opportunity statement]
- [N.2]: [Sub-opportunity statement]

**Dependencies**: [Does this opportunity depend on or enable other opportunities?]

Opportunity Tree Summary

After listing all opportunities, provide a tree visualization:

## Opportunity Tree

[Outcome / Root Opportunity]
├── [Branch Opportunity 1]
│   ├── [Leaf 1.1] ← actionable
│   └── [Leaf 1.2] ← actionable
├── [Branch Opportunity 2]
│   ├── [Leaf 2.1] ← actionable
│   └── [Leaf 2.2] ← actionable
└── [Leaf Opportunity 3] ← actionable

Task 5: Develop Recommendations (`recommendations.md`)

Tie specific actions to findings and opportunities:

# Recommendations

## Recommendation [N]: [Specific, actionable statement]

**Addresses**: Opportunity [X] / Finding [Y]

**What to do**: [Concrete action — specific enough to start on]

**Expected impact**: [What changes for users if this is done]

**Evidence basis**: [Brief reference to supporting data]

**Priority**: [From framework]

**Open considerations**: [Unknowns, risks, or tradeoffs to be aware of]

Task 6: Document Open Questions (`open-questions.md`)

Honest accounting of what the research did NOT answer:

# Open Questions & Future Research

## Unresolved Questions
- [Question]: [What we know so far + what's still unclear + why it matters]

## Suggested Follow-Up Research
- [Method]: [What question it would answer + why this method is appropriate]

## Limitations of This Analysis
- [Limitation]: [How it constrains what we can claim]

Task 7: Write Phase 4 Summary (`phase4-summary.md`)

# Phase 4 Summary: Synthesis Complete

## Personas: [Count] [Type] personas
[1-line description of each]

## Top Findings (ranked by priority):
1. [Finding] — [Prevalence] — [Confidence]
2. ...

## Opportunity Areas: [Count]
[1-line each]

## Recommendations: [Count]
[1-line each]

## Evidence Strength
[Overall assessment: where is evidence strong vs. thin?]

## Key Limitations
[Top 2-3 limitations the report must acknowledge]

## Ready for Report Compilation
[Any notes for Phase 5 — emphasis, audience considerations, structural suggestions]

Quality Gate

Before writing your outputs, verify:

Cognitive Empathy: Do personas read as real people with genuine perspectives, not cardboard cutouts? Would a participant recognize themselves?
Groundedness: Every finding is traceable to specific codes, themes, and quotes. No finding is pure analyst invention.
Minimum evidence: Every finding appears in at least 2 interviews. Remove any that don't meet this threshold or demote to "open questions."
Palpability: Every finding has at least 1 luminous exemplar quote and 2+ echo quotes. Evidence is concrete, not abstract.
Heterogeneity: Findings acknowledge variation and negative cases, not just the dominant pattern. Personas represent distinct groups, not slight variations on the same type.
Reflexivity: Prioritization framework choice is justified. Confidence levels are honest. Limitations are real, not pro-forma.
Outcome orientation: Opportunities describe desired outcomes, not just problems. Recommendations are specific enough to act on.

Parallel Extensibility Slot

The parallel/ directory is reserved for future analysis signals embedded in this phase. Currently empty. Examples of what could be added here:

journey-map.md — Synthesized journey stages across personas with emotional arcs
competitive-gaps.md — How findings compare to competitor approaches
severity-matrix.md — Severity ratings per finding using a standardized scale

On this page