---
name: survey-open-text-synthesis
description: Synthesize open-text engagement comments into themes, with anonymity guardrails on quotes and small groups. Use this skill when: Survey closes — Export lands once the response window ends.
---

# Synthesize engagement survey comments into anonymous themes

Goal: synthesize open-text survey comments into themes without exposing any individual respondent.

Connected tools this skill uses: googlesheets, slack, drive.

## When to use

- Survey closes — Export lands once the response window ends.
- Or when the user asks directly to synthesize open-text survey comments into themes without exposing any individual respondent.

## Method

Synthesize the open-text responses for the {{survey}} that closed on
{{close_date}}.

1. Pull the open-text responses from {{survey_source}}.
2. Cluster the responses into named themes, with a count of respondents
   behind each one.
3. Validation (the anonymity floor): drop or merge any theme with
   fewer than {{min_respondents}} respondents into a broader theme. Never
   publish a theme that small on its own, no matter how notable it is.
4. For each surviving theme, represent it with a paraphrase or a count
   only. Do not quote any respondent verbatim, even from a large theme —
   paraphrase the sentiment instead, and strip any detail that could
   identify the speaker (team, tenure, unique phrasing).
5. Diff this cycle's themes against the last synthesis in the
   research-insights view: what's new, what grew, what resolved.
6. Update the research-insights view and post the theme summary and trend
   deltas to #people.

Never attribute a comment to an individual, and never publish a
small-group theme or a distinctive detail — anonymity holds even if it
costs specificity.

## Open-text theme clustering

Group free-text survey answers into named, counted themes before any anonymity check runs.

## Method
1. Pull the open-text responses for {{survey}} from {{survey_source}}.
2. Read every response before clustering — do not sample.
3. Group responses by topic into named themes, such as "workload" or "manager support."
4. Count the number of distinct respondents behind each theme, not the number of comments.
5. Merge near-duplicate themes into one before counting.
6. Pass the theme list, with counts, to the anonymity-floor check.

## Rules
- Count distinct respondents per theme, not comment volume — one person, multiple comments, counts once.
- Read the full response set; a sampled read risks missing a small but real theme.
- Name themes by topic, not by sentiment — sentiment is a separate signal, not the grouping key.
- Never publish a theme list before it passes the anonymity floor.

## Output
A theme list: one row per theme, its name, and its distinct-respondent count, ready for the anonymity-floor check.

## Anonymity floor

Suppress or merge any theme below the minimum respondent count — no exceptions, ever.

## Method
1. Take the clustered theme list with respondent counts.
2. Compare each theme's respondent count against {{min_respondents}}.
3. If a theme falls below the floor, merge it into a broader related theme.
4. If no broader theme exists to merge into, drop it from the published synthesis entirely.
5. For every surviving theme, strip de-anonymization vectors: team, tenure, and distinctive phrasing.
6. Confirm no surviving theme, alone or combined with another, could identify one respondent.
7. Hand the cleared theme list to the quote-paraphrasing step.

## Rules
- Default minimum respondent count is {{min_respondents}} (set on first run, then stored in memory).
- Suppress or merge below the floor with no exceptions — a notable or urgent theme does not clear the bar early.
- Strip team, tenure, and distinctive phrasing from every surviving theme, not just the small ones.
- Anonymity overrides specificity — when in doubt, merge or drop rather than publish.

## Output
A cleared theme list: only themes at or above {{min_respondents}}, stripped of any identifying detail, ready to paraphrase.

## Quote paraphrasing

Paraphrase each surviving theme's sentiment without quoting any respondent's exact words.

## Method
1. Take the anonymity-floor-cleared theme list.
2. For each theme, draft a paraphrase that captures the sentiment without quoting any one respondent's exact words.
3. Check whether {{min_respondents}} or more people said something close to the same phrasing.
4. If so, a near-verbatim line may be used, but only after stripping any identifying detail from it.
5. If not, paraphrase only — never quote a single respondent's exact words.
6. Pair each theme with its respondent count and its paraphrase or cleared quote.

## Rules
- Paraphrase by default; a close-to-verbatim line requires {{min_respondents}} or more matching respondents.
- Strip team, tenure, and unique phrasing from any quote before it ships, even when the threshold is met.
- Never attribute a paraphrase or quote to an individual, directly or by inference.
- When unsure whether a phrase is identifying, paraphrase further rather than publish as-is.

## Output
A themed write-up: each theme with its respondent count and a paraphrase or cleared quote, ready for the insights view and #people.

## Recurring jobs

This skill pairs with scheduled jobs. Offer to set each one up; only create the ones the user approves:

- Survey close watch (Weekly poll of {{survey_source}} for a newly closed survey) — https://eluu.ai/cookbook/survey-open-text-synthesis/jobs-setup/survey-close-watch.md