
Description
What this scout watches and the condition under which it files a report.
SKILL.md
Your scout's title
One line on what it watches. Everything below is the agent's instructions, executed on every run, so write it in the second person to the agent – not to the reader.
Say how this relates to the canonical fleet. Most scouts you write will be a narrower
version of an existing one, and saying which saves a reader from running both. Check
products/signals/skills/ in the monorepo before you assume yours is new.
If your schedule in index.mdx isn't the fleet default of daily, justify it here in one
line.
Discriminator
The rule that decides whether a run files a report or writes nothing. Be specific enough that two different agents would reach the same verdict on the same data.
Prefer a rate over a raw count. A threshold on a count always needs a human to ask "but did traffic move?", which is why those alerts get muted. A threshold on a normalized rate answers that question itself – which is what lets the agent decide alone. Pick your denominator before you pick your threshold.
The bar
Give a concrete starting threshold, and say it's meant to be tuned:
- the numeric condition, on a complete bucket – a partial bucket reads as a spike every time, and
- a volume floor, below which the number is noise.
The volume floor usually matters more than the threshold.
Explore
Numbered steps the agent works through. Order them by likelihood, cheapest first.
- Rule out the instrumentation. Broken tracking looks identical to a broken product in almost every scout. Check this before anything else.
- Rule out the innocent explanation. Whatever your discriminator's quiet case is, this is where the agent confirms it doesn't apply.
- Find the cause. A report that says a number moved, with no cause attached, is something the reader still has to go investigate themselves.
- Tie it to a deploy where one sits in the window, and include the commit range.
Disqualifiers
The explicit list of "don't file for this". Every false positive you can think of goes here – this list is what keeps the inbox worth reading.
- The innocent explanation from your discriminator's quiet case.
- Anything deliberate and announced.
- Anything below the volume floor.
- Not enough history to have a stable norm yet.
More skills from the posthog.com repository
View all 4 skillssignals-scout-ai-cost
monitor AI feature costs and performance
Aug 10Cost OptimizationLLMMonitoringObservability +1signals-scout-flag-debt
identify and remove stale feature flags
Aug 10Feature FlagsObservabilityPostHogTech Debtsignals-scout-silent-failure
detect and diagnose silent application failures
Aug 10DebuggingMonitoringObservabilityPostHog
More from PostHog
View publisheranalyzing-expensive-users
analyze expensive users in AI observability
posthog
Aug 10AnalyticsCost OptimizationObservabilityPostHogauditing-endpoints
audit PostHog project endpoints
posthog
Jun 8AnalyticsAuditPostHogauditing-warehouse-source-health
audit PostHog data warehouse source health
posthog
Jun 18AuditData WarehouseObservabilityPostHogauditing-warehouse-view-health
audit PostHog materialized view health
posthog
Jun 18AuditData WarehousePerformancePostHogauthoring-error-tracking-alerts
author PostHog error tracking alerts
posthog
Jun 18AlertingDebuggingObservabilityPostHogauthoring-log-alerts
author log alerts in PostHog
posthog
Jul 18AnalyticsMonitoringObservabilityOperations +1