El Analizador de Resultados de Encuestas
Why this prompt matters
Teams that manually code open-ended survey responses spend an average of 2-3 full workdays per survey cycle doing it by hand, and rushed manual analysis under deadline pressure frequently misses early attrition signals buried in a handful of written comments — the same signals that show up as a resignation spike two quarters later once it's too late to act.
What we use it for
A People Ops manager just closed a 340-response employee engagement survey across the engineering organization and has 48 hours to present findings to the VP of Engineering before a headcount planning meeting, with only a raw CSV export and no built-in analytics dashboard.
Prompt
<p>Act as a senior people-insights analyst who has spent a decade turning raw survey exports into findings that change how leadership makes decisions.</p><p><strong>Context:</strong> I ran a survey titled [SURVEY NAME/PURPOSE — e.g. "Q3 Employee Engagement Survey"] with [NUMBER] respondents out of [TOTAL POPULATION] invited (response rate). The survey included both closed-ended (rating scale, multiple choice) and open-ended questions. The audience for your analysis is [AUDIENCE — e.g. "the VP of People and the executive team"], and the decision they need to make is [DECISION — e.g. "whether to fund a retention initiative for the engineering org"].</p><p><strong>Task:</strong> Analyze the raw survey data below and produce a findings report that helps the audience make that decision. Do not just describe what the data says — identify what matters most and what to do about it.</p><p><strong>Constraints:</strong><br/>- Report exact percentages and counts, never vague terms like "most" or "many."<br/>- Explicitly flag any finding based on a subgroup smaller than [MINIMUM SAMPLE SIZE — e.g. 15] respondents as low-confidence.<br/>- Distinguish correlation from causation — if two data points move together, say so, but do not claim one caused the other unless the data supports it.<br/>- When coding open-ended responses into themes, report how many distinct responses support each theme, and include 1-2 verbatim quotes (anonymized) per theme.<br/>- If closed-ended and open-ended responses conflict on the same topic, surface the conflict explicitly rather than picking one to report.</p><p><strong>Output Format:</strong><br/>1. <strong>Executive Summary</strong> (2-3 sentences: the single most important finding and its business implication)<br/>2. <strong>Quantitative Breakdown</strong> (key metrics by question, with response rate and any notable subgroup differences)<br/>3. <strong>Qualitative Themes</strong> (3-5 themes from open-ended responses, ranked by frequency, each with supporting quote(s))<br/>4. <strong>Sentiment Analysis</strong> (overall tone breakdown: positive/neutral/negative, and what's driving negative sentiment specifically)<br/>5. <strong>Recommended Actions</strong> (2-4 specific actions, each tied to a finding above, with a suggested owner)</p><p>[PASTE RAW SURVEY DATA HERE — CSV, table, or list of responses]</p>
Result
Executive Summary
Engineering engagement scores dropped from 7.8 to 6.4 (out of 10) this quarter, driven primarily by a sharp decline in perceived career growth (42% rated it 3/10 or lower, up from 18% last quarter) — this is the strongest single predictor of the resignation intent flagged in open-ended responses.
Quantitative Breakdown
- Response rate: 340/410 (82.9%), a strong sample.
- Overall engagement: 6.4/10 average, down from 7.8/10 last quarter (n=340).
- Career growth satisfaction: 42% rated 3/10 or lower (143 respondents), up from 18% (61 respondents) last quarter.
- Manager relationship score held steady at 7.6/10 — this is not where the decline originates.
- Note: the 12-person Platform sub-team scored notably lower (5.1/10 average) but this is a low-confidence finding given the small subgroup size.
Qualitative Themes
1. Lack of promotion clarity (61 mentions) — "I don't know what I need to do to get to senior, and neither does my manager."
2. Increased on-call burden (44 mentions) — "On-call has doubled since the reorg and nobody adjusted expectations."
3. Positive team culture (38 mentions) — "My immediate team is the best part of this job," consistent with the steady manager relationship score.
Sentiment Analysis
58% negative, 27% neutral, 15% positive on open-ended responses overall — a reversal from last quarter's 45% positive. Negative sentiment concentrates almost entirely around career progression and on-call load; sentiment toward direct managers and team culture remains positive, indicating the issue is structural/organizational rather than a management-quality problem.
Recommended Actions
1. Publish clear leveling criteria for senior promotion — People Ops + Eng Leadership, next 30 days.
2. Audit on-call rotation load post-reorg and rebalance — Eng Managers, next 2 weeks.
3. Follow up individually with the 12-person Platform sub-team given its lower score, before drawing org-wide conclusions from that group.
La mayoría de los prompts "analiza esta encuesta" devuelven una puntuación media y un párrafo vago sobre temas. Eso no es útil cuando la decisión sobre la mesa es financiar un programa de retención o reestructurar un equipo. El Analizador de Resultados de Encuestas está diseñado para sacar a la luz lo que realmente cambia una decisión, y para ser honesto sobre lo que los datos no pueden decirte.
Por qué este prompt está estructurado así
La sección de Constraints es la que hace el trabajo más importante. Exigir porcentajes exactos en lugar de "la mayoría de los encuestados" obliga al modelo a interactuar realmente con los números en lugar de parafrasearlos. La instrucción de marcar cualquier subgrupo por debajo de un tamaño de muestra mínimo como de baja confianza existe porque es extremadamente fácil para una IA — o un analista humano bajo presión de plazo — sacar una conclusión segura a partir de 8 personas y presentarla con la misma autoridad que un hallazgo basado en 300.
La restricción de correlación versus causalidad importa más en el análisis generado por IA que en el análisis humano, porque los modelos de lenguaje son lo suficientemente fluidos como para afirmar una relación causal de manera persuasiva incluso cuando los datos solo respaldan una correlación. Exigir al modelo que distinga entre ambas, y que saque a la luz conflictos directos entre las calificaciones de preguntas cerradas y los comentarios abiertos, en lugar de elegir silenciosamente la que respalde una narrativa más limpia, mantiene la honestidad del resultado.
Las cinco secciones del Output Format se corresponden directamente con lo que realmente necesita un tomador de decisiones. "Sentiment Analysis" pregunta específicamente qué está impulsando el sentimiento negativo — no solo el porcentaje — porque una puntuación negativa sin un factor identificado no es procesable. "Recommended Actions" exige que cada acción se vincule a un hallazgo específico anterior, lo que evita que el modelo genere consejos genéricos ('mejorar la comunicación') desconectados de lo que realmente mostró la encuesta.
Cómo usarlo bien
Rellena los campos de audiencia y decisión de forma específica — un análisis escrito para un VP que decide sobre un presupuesto de retención se ve diferente de uno escrito para un equipo de RRHH que diseña las preguntas de la encuesta del próximo trimestre. Para ciclos de encuestas recurrentes (compromiso trimestral, feedback posterior a eventos, NPS de clientes), guarda tu versión rellenada e intercambia los nuevos datos brutos cada ciclo para obtener una estructura de informe consistente que puedas comparar trimestre a trimestre.
Conclusiones
El valor aquí no está en comprimir datos — está en forzar la misma disciplina que aplicaría un analista humano cuidadoso: distinguir lo que es estadísticamente sólido de lo que es una señal débil, rastrear el sentimiento hasta su factor real, y vincular cada recomendación a evidencia específica. Úsalo siempre que una decisión dependa de datos de encuestas y no tengas días para codificar manualmente las respuestas abiertas.