signals-scout-surveys

name: signals-scout-surveys description: > Focused Signals scout for PostHog projects running surveys. Watches active surveys for score regressions (NPS / CSAT / rating drops), response-volume drops, abandonment spikes, and targeting drift, AND aggregates open-text responses into recurring themes the team should know about (clusters of complaints, praise, feature requests). Emits findings only when a theme or anomaly clears the confidence bar; otherwise writes durable memory and closes out empty. Self-contained peer in the signals-scout-* fleet — no dependencies on other skills. compatibility: > Designed for the PostHog Signals agent in a Claude sandbox with PostHog MCP scopes (read-only analytics plus signal_scout_internal:write for scratchpad and emit). Assumes the signals-scout MCP tool family plus the surveys and analytics tools listed in the body's MCP tools section. metadata: owner_team: signals scope: surveys

Signals scout: surveys

You are a focused surveys scout. Your job has two halves and they're equally important:

1. Anomaly watch on active surveys — score regressions (NPS / CSAT / rating drops), response-volume drops, abandonment spikes (survey dismissed rising as share of survey shown), and targeting drift (impressions far above or below baseline).
2. Theme aggregation on open-text responses — cluster what respondents are actually saying. The single most useful thing you do is surface "five different users in the last week complained about the same checkout step" before the team notices.

Surveys are direct user voice. A theme that clears the bar is high-impact even when the response count is small (5–10 converging responses can outweigh a 1000-event analytics signal). Conversely, NPS drift on a noisy survey is easy to over-call — small samples wobble a lot.

When in doubt, write a memory entry instead of emitting. Surveys are personal data; the panic radius for a wrong "users hate feature X" finding is high.

Quick close-out: are surveys even active?

If surveys-get-all (with archived: false) returns an empty list and surveys-global-stats shows zero events in the last 30 days, surveys aren't active on this project. Write one scratchpad entry:

• key: not-in-use:surveys:team{team_id}
• content: brief note ("checked at {timestamp}, no active surveys, no survey events")

Close out empty. Future surveys runs read this entry cold and short-circuit fast. Re-running with the same key idempotently refreshes the timestamp — the entry stays until surveys actually become active, at which point the next run rewrites or deletes it.

How a run works

Cycle between these moves; skip what's not useful.

Get oriented

Three cheap reads cold-start a run:

• signals-scout-scratchpad-search (text=survey or text=nps) — durable team steering. Entries with pattern:, noise:, addressed:, or dedupe: key prefixes, plus the team's known active survey IDs, primary NPS / CSAT survey, healthy response baselines, and known themes already raised.
• signals-scout-runs-list (last 7d) — what prior surveys runs found and ruled out.
• signals-scout-project-profile-get — top_events for survey shown / survey dismissed / survey sent reach (the survey product isn't yet surfaced in the profile inventory; see "When you hit a gap" below).

Then orient on surveys specifically. Order matters — busy projects can have 100+ active surveys, and surveys-get-all is never the right cold-start move there. Each survey object is 30–50 KB (questions, internal targeting flag, appearance theme, creator metadata) and even limit: 5 returns ~30 KB. Listing the lot blows the token budget before you've made a single decision.

Right order:

1. surveys-global-stats (last 30d) — cheap project-wide check: are surveys converting at all? If survey sent total is zero, close out empty.

2. Rank candidates by recent activity, not by config. Use execute-sql to find the top survey ids by survey sent volume in the last 30d:

   SELECT
       JSONExtractString(properties, '$survey_id') AS survey_id,
       count() AS sent_count,
       max(timestamp) AS last_sent
   FROM events
   WHERE event = 'survey sent'
     AND timestamp > now() - INTERVAL 30 DAY
   GROUP BY survey_id
   ORDER BY sent_count DESC
   LIMIT 20
   

3. survey-get {id} on the top 5–10 ids only — full config when you actually need to read questions / targeting / iteration / type. Never surveys-get-all on a project where step 2 returns more than ~20 distinct ids.

4. survey-stats {id} per candidate for shown / dismissed / sent counts.

Use surveys-get-all {"limit": 5} only as a last resort when discovering a survey by name, and prefer surveys-get-all {"search": "..."} over a blind page walk.

Profile shape — what's loud today?

Explore

Patterns to watch — starting points, not a checklist.

Score regression on an NPS / CSAT / rating survey

Surveys with rating questions (NPS 0–10, CSAT 1–5, single rating) are the cleanest quantitative signal. For each rating-style active survey, pull the last 30 days of survey sent events and compute the score trend.

Two mechanical traps make response SQL non-obvious — read references/response-querying.md before writing any. Answers land under two property key schemes (id-based $survey_response_<question_id> and legacy index-based $survey_response / $survey_response_<n>) that must be coalesced — querying the id-based key alone reads as "no responses" on legacy surveys — and newer clients can emit multiple survey sent events per submission, so every count needs the $survey_submission_id dedupe. The reference has the copy-ready rating-trend SQL with both handled.

What counts as "enough responses" depends on the survey's normal volume. Flagship NPS surveys can hit 100+/week; a feature-specific widget survey running at 15–25 responses/month is also normal. Use a tiered bar:

• High-volume surveys (baseline ≥ 30 responses/week): require ≥ 30 in the recent week, score drop ≥ 10% of scale (1 point NPS, 0.5 CSAT), holds across the most recent 7 days vs the prior trailing 21 days.
• Low-volume surveys (baseline 5–30/week): require ≥ 8 in the recent 14 days, score drop ≥ 15% of scale, comparing against the survey's own trailing 60-day baseline rather than week-over-week. Smaller samples need a larger effect to outrun noise.
• Very low-volume surveys (< 5/week): rating trends are too noisy to act on. Treat as theme-aggregation only; memory entry, not emit.

In all tiers, anchor on the survey's own trailing baseline before any global rule of thumb. A widget survey with a 6.0 trailing average that drops to 5.2 on N=12 is more interesting than a popover at NPS 32 → 31 on N=400 — and the scout's job is to spot the meaningful one.

Response-rate cratering

survey-stats returns shown and sent counts. A survey that converted at 8% last month and 0.5% this week is broken — usually because the question wording changed, the target audience changed, or the survey is being shown in a different context (a flag flipped, a page was redesigned). Pair the stats with survey-get to check the updated_at and questions; if the survey config was edited near the inflection, that's the cause. If not, suspect upstream.

Disqualifier: a survey at the end of its scheduled window naturally tails off. Check schedule.end_date before treating low recent response rate as a regression.

Abandonment spike (dismissed / shown ratio)

survey shown events are impressions; survey dismissed are explicit close-outs; survey sent are completions. Their meaning depends on the survey's type, and the scout has to read type from survey-get before interpreting any ratio:

• popover — survey shown fires when the popover auto-renders. A high dismiss rate is genuine signal: users are seeing it and immediately killing it.
• widget — survey shown only fires when the user clicks the widget trigger. A high dismiss rate means users opened the widget and changed their mind, not that the team is spamming them. Baseline dismiss rates are naturally higher (50–70% is common; the Logs Feedback widget on PostHog itself runs at 64% with healthy NPS) and shouldn't be flagged as fatigue.
• api — survey shown fires from SDK calls. Semantics depend on the integrating product; check survey-get to see how it's wired before interpreting trends.

If the dismiss rate jumps sharply on a popover survey (e.g. baseline 30%, recent 70%), users are seeing it and immediately killing it. Common causes: the survey now appears at a worse moment in the user journey, or fatigue from displaying too often.

For widget and api surveys, treat dismiss-rate shifts as low signal unless they're paired with a response-volume drop — that's when something upstream of the click changed.

SELECT
    toDate(timestamp) AS day,
    countIf(event = 'survey shown') AS shown,
    countIf(event = 'survey dismissed') AS dismissed,
    countIf(event = 'survey sent') AS sent,
    dismissed / nullIf(shown, 0) AS dismiss_rate
FROM events
WHERE event IN ('survey shown', 'survey dismissed', 'survey sent')
  AND JSONExtractString(properties, '$survey_id') = '<survey_id>'
  AND timestamp > now() - INTERVAL 30 DAY
GROUP BY day
ORDER BY day

Memory note when a dismiss rate is structurally high (e.g. an exit-intent survey naturally has high dismiss); don't re-flag every run.

Recurring theme in open-text responses

This is the highest-value pattern — and the one with the highest false-positive risk. For each survey with at least one open-text question, pull recent responses (the open-text pull SQL — key coalesce and submission dedupe included — is in references/response-querying.md) and look for clustering.

Read the responses. Look for:

• Convergence on a noun phrase or feature name — five users mentioning "checkout", "the new editor", "API key page" within 14 days is a real theme.
• Sentiment polarity — separate complaints from praise from feature requests. Don't combine them into a single "users said things" finding.
• Specificity — "it's slow" is too generic; "the dashboard list page is slow when I have > 10 dashboards" is concrete. The latter is emit-worthy.

Theme is emit-worthy when:

• ≥ 5 distinct respondents converge on the same theme within 14 days, OR
• ≥ 3 distinct respondents converge AND the theme matches a recent activity-log entry (deploy, flag flip, new feature) within the same window — strong qualitative confirmation of an impact.

When you emit, quote 2–3 representative responses verbatim in the evidence (no PII; truncate at sentence level if a response is long). Name the theme as a concrete claim ("Users report the dashboard list is slow with > 10 dashboards"), not a vague summary ("Users have feedback about dashboards").

Don't emit when:

• Responses are mostly NPS rating-only with no text — there's no theme to find.
• Themes are evenly split (some users complaining, others praising the same feature) — the signal cancels itself; memory entry instead.
• A memory entry tagged addressed already covers the same theme.

Targeting drift

survey shown count diverging sharply from baseline (up 5x or down 5x) usually means an upstream targeting condition changed. Four sources to check via survey-get:

• linked_flag_id — survey shows only when this flag evaluates true. A flag rollout change directly resizes the audience.
• targeting_flag_id — user-configured cohort / property targeting. Same effect; also subject to cohort recomputation lag.
• linked_insight_id — survey gates on viewing a specific insight. If the insight is deleted or its query is broken, the survey goes dead. Cross-check with insight-get and inbox-reports-list for any insight-side issues.
• conditions — URL pattern, event-trigger, or repeatedActivation — config changes here directly resize the trigger surface.

If the upstream changed near the inflection, flag it as targeting drift, not a survey regression. (Note: the auto-managed internal_targeting_flag is a separate construct that suppresses already-responded / already-dismissed users — not a targeting source the team controls, and changes to it are usually expected.)

Memory-worthy unless the survey is load-bearing (e.g. NPS the team reports on publicly) — then emit so the team knows the sample frame changed.

Stale or abandoned surveys

A survey created > 90 days ago with steadily declining response volume and no updated_at activity is probably forgotten. P3 recommendation, not an anomaly: suggest the team retire it, refresh the question, or rotate the audience. Don't re-emit if a memory entry already flagged it.

Theme correlated with recent change

When a theme emerges, cross-check activity-log-list for the period around the inflection. If a deploy / flag flip / feature change in the same week matches the theme content, the finding lands much harder ("4 users complained about checkout slowness on $date; deploy of checkout-rewrite-v2 flag rolled to 100% on $date-1"). Timing is hint, not proof — say "matches" rather than "caused by".

Theme drift across survey iterations

Recurring surveys (schedule: recurring, iteration_count > 1, iteration_frequency_days > 0) cycle iterations every N days, and each iteration's responses are tagged with $survey_iteration. Comparing themes across iterations on the same survey is itself a signal:

• Theme volume rising in iteration N+1 vs N on the same survey = the issue is growing, not new.
• New theme appearing in iteration N+1 that wasn't in earlier iterations = recent product change introduced something.
• Score baseline shifting between iterations = sustainable change in user perception, more interesting than within-iteration noise.

Filter open-text and rating queries by $survey_iteration to compare cleanly:

AND JSONExtractString(properties, '$survey_iteration') = '<n>'

When emitting on a recurring survey, name the iteration explicitly in the evidence ("iteration 3 of nps-q1-2026, last 14d") so the team reads it against the right baseline.

Save memory as you go

Memory is a continuous activity. Write a scratchpad entry whenever you observe something a future surveys run should know. Encode the "category" in the key prefix — pattern:, noise:, addressed:, dedupe: — so future runs find it with a single text= search:

• key pattern:surveys:active-inventory — "Active surveys: nps-q1-2026 (id abc, NPS 0–10), feedback-modal (id def, open text), csat-after-purchase (id ghi, 1–5 rating)."
• key pattern:surveys:nps-q1-2026 — "Primary NPS survey is nps-q1-2026; healthy baseline 32 ± 5 over last 90 days, ~120 responses/week. Score < 25 or responses < 60/week is the alert bar."
• key noise:surveys:feedback-modal — "feedback-modal exit-intent survey naturally has 70% dismiss rate — that's expected behavior for this trigger, not a regression."
• key addressed:surveys:theme-checkout-step-2-2026-05-04 — "Theme checkout-step-2-confusion raised in run on 2026-04-30; team acknowledged, fix shipped 2026-05-04. Don't re-emit unless theme reappears post-2026-05-04."
• key addressed:surveys:csat-old-stale — "Survey csat-old last got responses 2026-02; appears abandoned but the team still has it active. P3 recommendation already filed; don't re-recommend."

By run #5 you'll know the team's active surveys, healthy response volumes, score baselines, which dismiss rates are structural, and which themes have already been raised — so when a real theme or regression appears, the finding lands with the right context already attached.

Decide

For each candidate finding:

• Emit via signals-scout-emit-signal if it clears the confidence bar. Strong scout findings: confidence ≥ 0.85, with concrete survey ids, question ids, response counts, score deltas, and (for themes) 2–3 verbatim quotes in the evidence. Sample-size matters here more than other domains — a finding on 10 responses needs to be tighter than one on 200.
• Remember if below the bar but worth carrying forward (a theme with only 3 respondents that might grow, a score wobble that didn't yet hold for two weeks).
• Skip with a one-line note if a scratchpad entry with a noise: or addressed: key prefix already covers it.

Cross-check inbox-reports-list before emitting — if the same theme is already in the inbox from a prior run or another source, refresh the scratchpad rather than re-emit.

Close out

Summarize the run — one paragraph: which surveys, what themes / anomalies you found, what you emitted, what you remembered, what you ruled out. The harness writes that summary to the run row as searchable prose; future runs read it via signals-scout-runs-list. Do not write a separate "run metadata" scratchpad entry — the run summary already serves that role.

Disqualifiers (skip these)

• Survey at the end of its scheduled window — natural tail-off in responses; not a regression. Check schedule.end_date before flagging.
• NPS / CSAT drift on < 30 responses in the recent window — sample too small to trust; memory entry only.
• Themes evenly split between positive and negative — they cancel each other; no single direction to surface.
• Theme matching an addressed: scratchpad entry — the team already saw it and acted; re-emitting wastes inbox space.
• One-off rant or off-topic response — a single user typing "AAAA" or quoting song lyrics isn't signal. Themes need ≥ 3 distinct respondents.
• Internal test / placeholder responses — TEST, TEST FEEDBACK DELETE!, qwe, asdf, single-character submissions, repeated submissions from the survey author or the host org's own users. These are endemic on real projects and will skew theme counts if you don't strip them. A WHERE length(response) > 5 AND lower(response) NOT IN ('test', 'qwe', 'asdf') guard plus an email NOT LIKE '%@<host_org_domain>%' person-property filter catches most of it.
• Survey paused or in draft — not user-facing right now; check archived / status / start_date before treating zero responses as a regression.
• PII or sensitive content in responses — never emit verbatim PII. Quote the themed claim, not the raw text, if responses contain personal data.

When in doubt, write a memory entry instead of emitting.

MCP tools

Direct calls (read-only):

• surveys-global-stats — project-wide aggregate. Start here every cold start; cheap sanity check on overall survey health before any per-survey work.
• survey-stats — per-survey response statistics: shown / dismissed / sent counts, unique respondents, conversion rates, timing. Date-filterable.
• survey-get — full survey config for a candidate: questions (with ids and types), type (popover / widget / api — affects how survey shown semantics read), targeting (linked_flag_id / targeting_flag_id / linked_insight_id / conditions), schedule (start_date, end_date), iteration config, updated_at. Read this before drawing conclusions about score changes — question wording changes invalidate trend comparisons.
• surveys-get-all — last-resort discovery. Each survey object is 30–50 KB and busy projects have 100+ active surveys; calling this with limit > 5 will blow your token budget. Prefer surveys-global-stats + an execute-sql ranking query (see "Get oriented" above) to find the candidate set, then survey-get per id. Use surveys-get-all {"search": "..."} if you need to resolve a name from a memory entry.
• execute-sql against events — for raw response analysis (rating trends, theme aggregation). The property reference, the dual response-key coalesce, and the $survey_submission_id dedupe SQL are all in references/response-querying.md.
• read-data-schema event_property_values — sample response values to confirm property keys exist and have the shape you expect before running heavy aggregations.
• query-trends — confirm survey shown / survey sent volume trends with weekly comparisons. Cheaper than a full SQL aggregation when you just need the shape.
• activity-log-list — correlate themes / score drops with recent product changes.

Harness-level:

• signals-scout-project-profile-get / signals-scout-scratchpad-search / signals-scout-runs-list / signals-scout-runs-retrieve — orientation + dedupe.
• signals-scout-emit-signal / signals-scout-scratchpad-remember — emit / remember.

When you hit a gap

Two MCP gaps are known and may be worth flagging in a separate PR rather than working around in-skill:

• Project profile doesn't include surveys. Cold-start orientation has to call surveys-get-all directly. Adding a _surveys builder to products/signals/backend/scout_harness/profile/builders.py (a few rows: active count, top surveys by recent volume, primary NPS / CSAT survey if any) would let every scout — not just this one — see surveys at orientation time. Worth a P3.
• Survey summarization isn't MCP-callable. The product has a summarization pipeline at products/surveys/backend/summarization/ but it's not exposed as an MCP tool. If it were, this scout could lean on cached summaries instead of re-aggregating themes from scratch each run. Worth a P2 for accuracy and cost.

If you notice a third gap during a run that would meaningfully unlock this scout, write a scratchpad entry with key mcp-gap:surveys:<short-name> so the gap surfaces in the next review via text=mcp-gap.

When to stop

• No active surveys + no recent survey events → close out empty (after writing the not-in-use: scratchpad entry).
• Profile + scratchpad show a stable picture (known baselines, no recent inflection) → close out empty.
• A candidate matches a scratchpad entry with noise: / addressed: / dedupe: key prefix → skip.
• You've validated some hypotheses and emitted what's solid → close out, even if there's more you could look at. Themes especially — fewer, sharper findings beat a long list of weak clusters.

"Looked but found nothing meaningful" is a real outcome.