Could this accidentally re-identify someone?

That's the risk the whole system is built to prevent, and it's never zero — so you manage it deliberately. Set a floor in the first step (no theme, count, or quote tied to a group smaller than ~5), scrub every quote of names, team names, roles, and fingerprinting detail, and watch for rare roles and small segments where one person is the data. The biggest trap is combination: a safe-looking cut plus a telling quote plus a small team can unmask someone no single piece would. Read every quote by eye before it leaves the workspace — Claude flags, you decide.

Is it safe to put survey responses into Claude?

Only inside a Claude Desktop workspace your company has approved for people data, and you approve each read in the 'Ask permissions' prompt. Survey responses are among the most sensitive data HR holds — the whole exercise depends on respondents trusting their answers stay confidential, so the data never leaves the approved workspace and never gets pasted into an unapproved tool. If a workspace hasn't been cleared for people data, don't use it for this.

Does it just summarize, or does it actually prioritize?

It prioritizes. A summary would flatten every comment into one paragraph and let the loudest voice win. This ranks themes by how many distinct people raised them (prevalence, not word count), then sorts them by how widespread and how actionable they are — explicitly separating loud-but-rare from quiet-but-widespread — and ends in a top-3 action plan with owners and dates. The ranking is a recommendation you challenge, but it's a real prioritization, not a recap.

How is this different from the score dashboard our survey tool already gives us?

The dashboard tells you that engagement is 6.8 and dropped two points; it can't tell you why, because the why lives in the open-text comments it leaves as an unread column. This system reads those comments, clusters them into the handful of themes driving the scores, quotes them anonymously so leaders feel the issue, and turns the top ones into owned actions — plus an honest note on what you're choosing not to change. The dashboard is the symptom; this is the diagnosis and the plan.

Turn an engagement survey into an action plan

when to reach for this

The engagement survey closes and you get a CSV of free-text comments and scores that either becomes a generic "we hear you, we'll do better" slide or sits unread in a folder. The fix is to cluster the comments into the handful of themes that actually matter, separate loud-but-rare from quiet-but-widespread, keep the verbatim quotes that make leaders feel it — scrubbed of anything identifying — and turn the top themes into specific, owned actions. Anonymity is the whole reason people told you the truth: protect it or you poison next year's response rate. So this system is built backwards from one rule — no theme, no quote, no cut of the data is allowed to point a finger at a person who answered in confidence.

gather this first

The survey export — survey-comments.csv or similar, with the open-text answers and the scores — opened in a Claude Desktop workspace your company has approved for people data. Approve each read in the 'Ask permissions' prompt when it asks. This file is the trust contract; it does not leave the approved workspace.
The response counts — overall and per segment (team, tenure, location) — so you know up front which cuts of the data are too small to be safe to name. A theme tied to a three-person team is a theme tied to three people.
Last cycle's themes if you have them, so you can read trend instead of a snapshot — and a quick map of who owns which kind of action (managers, HR, leadership, facilities) so the action plan has real names on it.

the workflow

Set the anonymity rules before any analysis

Open the export in Claude Desktop and, before asking for a single theme, set the guardrails — what gets stripped and what's too small to name. Anonymity isn't a step at the end; it's the constraint everything else runs inside.

you ask

Read survey-comments.csv. Before any analysis, follow these rules for everything you produce: strip names, manager names, project names, and any specific detail that could identify who wrote a comment. Never tie a theme, a quote, or a count to any group smaller than 5 respondents — if a cut is that small, fold it up into a larger group or leave it out, and tell me you did. Confirm the row count and that you understand these rules before continuing.

what you get back A confirmation — "412 responses, here are the rules I'll hold to" — and a note that any segment under your floor will be merged or dropped. This is the contract for the whole session; everything downstream inherits it.

Set the floor explicitly and out loud. Re-identification is rarely one obvious name — it's a rare role plus a small team plus a telling detail, and you only catch that if the rule is stated before the analysis, not bolted on after.
Cluster the open-text into themes weighted by prevalence, not volume

The handful of people who write three paragraphs are not the majority. Ask for themes ranked by how many distinct people raised them, so a quiet, widespread concern beats a loud, lonely one.

you ask

Cluster the open-text comments into the 5–8 themes the survey is really about. Rank them by how many DIFFERENT respondents raised each one — prevalence, not word count. A long, detailed comment counts once, the same as a one-liner. For each theme give a respondent count, a percentage of responses, and a one-line description. Don't pull quotes or write actions yet.

what you get back A ranked list — "workload & pace (raised by 138, 33%), career growth (96, 23%), manager consistency (71, 17%)…" — each a real pattern with a number behind it, not a single vivid complaint dressed up as a trend.

Counts are what make this credible. "People are stressed about workload" moves no one; "a third of respondents raised workload" moves a leadership team — and keeps you honest about what's actually widespread.
Separate sentiment from signal

A theme can be loud and rare, or quiet and widespread — and they need different responses. Make Claude sort them apart explicitly, so a vocal minority doesn't quietly set the agenda.

you ask

Now sort the themes on two axes: how WIDESPREAD each is (how many people) and how ACTIONABLE it is (can we actually do something about it). Call out explicitly which themes are loud but rare versus quiet but widespread, and which are high-emotion but outside our control. Rank by widespread AND actionable together, and explain any case where the two pull in different directions.

what you get back A two-axis read — "career growth: widespread and very actionable → top priority; office coffee: loud in a few comments but rare and minor; pay vs. market: widespread but slow to act on, flag for leadership" — with the loud-and-rare cases named so they don't crowd out the quiet-and-real ones.

This is the discipline of the whole system. The vivid, angry, three-paragraph comment is the one that feels urgent — and is exactly the one prevalence-weighting is there to keep in proportion.
Pull representative, anonymized quotes per theme

A leadership team acts on what they feel, and a real verbatim line lands harder than any count. But every quote is a re-identification risk, so this step is half quote-pulling, half scrubbing.

you ask

For each top theme, pull 2–3 short verbatim quotes that capture it honestly — including the uncomfortable ones, not just the polite ones. Then scrub every quote: remove names, manager names, team names, project names, locations, job titles, and any specific detail that could fingerprint the person. If a quote can only be cleaned by gutting its meaning, drop it and pick another. Flag any quote you're not fully confident is anonymous.

what you get back A handful of scrubbed, representative quotes per theme — the lines that make the issue real — each safe to read aloud in a room, with anything borderline flagged for you to check by eye before it goes anywhere.

Read every quote yourself before it leaves the workspace. "My manager on the data team" plus a small team is a name, even with no name in it. You are the last line; Claude flags, you decide.
Draft the action plan — and what you're NOT doing

Top themes become specific, owned, dated actions. But the honest other half is naming what you heard and chose not to change — because a survey that pretends to act on everything is the one nobody believes next year.

you ask

Turn the top 3 themes into an action plan: for each, a specific action (not "we'll focus on wellbeing" — an actual thing that happens), a named owner, and a date. Then add an honest "what we heard and are NOT changing right now" section: 1–2 themes we're choosing not to act on, with the real reason why. Keep it plain and non-defensive. End with a one-line summary I can put at the top of an all-hands readout.

what you get back A tight plan — three owned, dated actions plus an honest "here's what we're not doing and why" — that people can hold you to. The not-doing section is what earns the trust, not the polished slide.

"We heard the request for a four-day week and we're not doing it this year, here's why" is a real, respectable output. Honesty beats a PR slide every time, and people can tell the difference.

make it your own

Read the trend, not just the snapshot: feed last cycle's themes alongside this one and ask what moved, what didn't, and what's new — "workload dropped from 41% to 33%, growth is up and now the top concern." A direction of travel lands harder with leadership than a single-year number, and it shows last year's actions actually did (or didn't) something.
Segment with extreme care: cutting by team or tenure is where re-identification happens. Hold the floor — never report a segment below your minimum, watch for rare roles (the one person in a function), and remember that several safe-looking cuts can combine to unmask someone. When in doubt, report the company-wide theme and skip the slice.
Hand off to the all-hands readout: pair this with Draft a sensitive message you can stand behind — the themes and the honest not-doing list become the readout that lands as candid rather than corporate. The action plan is the what; that playbook governs how you say it to a room that's watching whether you mean it.
Work in Arabic where that's the language: if your people answered in Arabic, analyze and present in Arabic — sentiment is carried in tone, idiom, and register that don't survive a round-trip through English, and a machine-translated theme can flip a wry complaint into a neutral one or miss the politeness that softens a real grievance. Run a respectful exit and keep what you learn is another signal source worth reading in the same language; exit themes and survey themes usually rhyme.
Save the flow as a `/pulse` skill (Power Track): once the prompts are stable, turn them into a reusable /pulse skill so every cycle runs the same anonymity rules and the same structure. Skills are the opt-in Power Track — on Desktop you keep these prompts in a shared note and run them in order each cycle, which works exactly as well until you're ready to automate.

watch out for

Re-identification is the cardinal risk. Small segments, rare roles, and a too-specific verbatim detail can each unmask a respondent — and combined, they're worse. Hold the floor below which you never name a group, scrub every quote, and read each one by eye before it leaves the workspace. One unmasked respondent and you've taught everyone not to be honest next year.
People data stays in an approved workspace. Survey responses are some of the most sensitive data HR holds. Keep the CSV and the analysis inside a Claude Desktop workspace your company has cleared for people data, approve each read in the 'Ask permissions' prompt, and never paste raw comments into an unapproved tool to "just check something."
Loud is not widespread. A vocal minority writes the longest, angriest comments and feels like the whole company — that's exactly what prevalence-weighting corrects for. And don't cherry-pick the comfortable themes either: the point is the truth the data tells, not the version that's easiest to present upward.
Claude finds the themes; leaders own the actions. The clustering and ranking are recommendations to challenge, not verdicts — present them as "here's what the data suggests," not "here's what's wrong." And "we decided not to act on X, and here's why" is a real, honest output, not a failure of the process.

you'll end up with A survey turned from an unread CSV into the few themes that actually matter — ranked by real prevalence rather than volume, separated into loud-but-rare versus quiet-but-widespread, quoted in anonymized verbatim lines that make leaders feel it, and shaped into an owned, dated action plan (including an honest list of what you're choosing not to change) that people can hold you to — all without ever putting a name to a respondent who answered in confidence.

Questions people ask

Could this accidentally re-identify someone?: That's the risk the whole system is built to prevent, and it's never zero — so you manage it deliberately. Set a floor in the first step (no theme, count, or quote tied to a group smaller than ~5), scrub every quote of names, team names, roles, and fingerprinting detail, and watch for rare roles and small segments where one person *is* the data. The biggest trap is combination: a safe-looking cut plus a telling quote plus a small team can unmask someone no single piece would. Read every quote by eye before it leaves the workspace — Claude flags, you decide.
Is it safe to put survey responses into Claude?: Only inside a Claude Desktop workspace your company has approved for people data, and you approve each read in the 'Ask permissions' prompt. Survey responses are among the most sensitive data HR holds — the whole exercise depends on respondents trusting their answers stay confidential, so the data never leaves the approved workspace and never gets pasted into an unapproved tool. If a workspace hasn't been cleared for people data, don't use it for this.
Does it just summarize, or does it actually prioritize?: It prioritizes. A summary would flatten every comment into one paragraph and let the loudest voice win. This ranks themes by how many distinct people raised them (prevalence, not word count), then sorts them by how widespread *and* how actionable they are — explicitly separating loud-but-rare from quiet-but-widespread — and ends in a top-3 action plan with owners and dates. The ranking is a recommendation you challenge, but it's a real prioritization, not a recap.
How is this different from the score dashboard our survey tool already gives us?: The dashboard tells you *that* engagement is 6.8 and dropped two points; it can't tell you *why*, because the why lives in the open-text comments it leaves as an unread column. This system reads those comments, clusters them into the handful of themes driving the scores, quotes them anonymously so leaders feel the issue, and turns the top ones into owned actions — plus an honest note on what you're choosing not to change. The dashboard is the symptom; this is the diagnosis and the plan.