Exit Interview Synthesizer

---
name: exit-interview-synthesizer
description: Use when synthesizing multiple exit interview notes into themes and recommended actions. Triggers on "synthesize these exit interviews", "exit interview themes", "patterns from departures".
---

# Exit Interview Synthesizer

Exit interview data is high-signal but easy to misread. One person's complaint isn't a theme. Five people raising the same issue from different angles IS. This skill separates them.

## Required inputs

1. **The notes** — multiple exit interviews (minimum 3, ideally 5+)
2. **Department / team / level** of the people leaving (helps stratify themes)
3. **Time window**
4. **Anonymity constraints** — can themes cite specific examples or do they need to be aggregated?

## Synthesis framework

### Step 1: Extract claims
Read every note. For each interview, list 3-7 specific claims as bullet points. Tag each:
- **[CULTURAL]** — environment, behavior, leadership style
- **[STRUCTURAL]** — comp, hierarchy, role definition, process
- **[RELATIONAL]** — specific relationships (manager, peer, leader)
- **[CAREER]** — growth, development, lack of progression
- **[OPERATIONAL]** — tools, workload, on-call, hours

### Step 2: Cluster
Group similar claims. Patterns emerge when:
- 3+ unrelated departures cite the same thing
- The same SPECIFIC person/process/team is named
- The same UNDERLYING issue is described in different language

Single mentions are NOT themes. Note them separately as "isolated."

### Step 3: Signal vs noise

For each potential theme, ask:
- **Source diversity**: Are these from different teams, levels, tenures? Or one cluster?
- **Specificity**: Are they specific enough to act on? Or vague ("management was bad")?
- **Recency**: Is this a current issue or about something already addressed?
- **Departure context**: People leaving for compelling new jobs report differently than people pushed out — note both.

### Step 4: Themes

Present 3-7 themes in priority order. For each:
- **Theme name**: short, descriptive
- **What people said** (paraphrased, not quoted)
- **Number of mentions** out of total interviews
- **Specificity / actionability**: high / medium / low
- **What the theme might be evidence of** (with humility — theme ≠ root cause)

### Step 5: Isolated signals

Single-source claims that didn't reach theme threshold but might be:
- Early warning of an emerging issue
- Specific to a person/process worth watching
- High-severity even if N=1 (e.g. harassment claim)

### Step 6: Recommended actions

3-5 specific actions, mapped to themes. Each:
- **Action**: concrete
- **Owner**: who would drive
- **Why now**: which theme(s) it addresses
- **Validation**: how you'd know it worked (next quarterly exit review, pulse survey, etc.)

Avoid generic recommendations:
- BAD: "Improve manager training"
- GOOD: "Manager X has been named in 3 of 5 exits in their org. Recommend an external 360 within the quarter."

## Bias controls

- **Don't make every theme about leadership** — sometimes leadership is the issue, sometimes the obvious narrative is wrong
- **Don't average out specific complaints** — "people felt disrespected by leadership" loses signal compared to "3 ICs reported being talked over by VP X in design reviews"
- **Survivor bias warning**: people who LEAVE may not represent people who stay
- **Don't anonymize past the point of usefulness** — but flag where naming specific people is appropriate vs not

## Anti-patterns

- "Everyone said comp could be better" — uselessly generic
- "Themes were leadership, growth, and culture" — three meaningless buckets
- Treating one impassioned complaint as a theme
- Treating a quiet recurring complaint as noise

## Output format

Markdown with clear sections:
1. Executive summary (3-5 sentences)
2. Methodology (1 paragraph: how many interviews, time window, stratification)
3. Themes (priority order)
4. Isolated signals
5. Recommended actions
6. What this synthesis CANNOT tell you (limits of the data)