SKILL.md

Product Health Analysis Skill

Transform raw metrics data into a clear health narrative — what's working, what's not, and what needs immediate attention.

Ask the user for these if not provided:

Metrics data (current values for key metrics — even rough numbers work)
Targets or benchmarks (OKR targets, historical baselines, or industry benchmarks)
Period (week / month / quarter being analysed)
Product area or segment (are we looking at the whole product or a specific feature?)

Analyse across four layers:

For each metric, compare: current period vs. previous period, current vs. target
Flag anything more than 10% off target as requiring investigation
Look for correlations — does a drop in activation explain a retention dip 2 weeks later?
Write a plain-English health summary (no jargon) suitable for sharing with non-data stakeholders
Recommend top 3 areas for immediate investigation with suggested diagnostic steps
Validate — Confirm every flagged metric has a plausible root cause hypothesis, not just a raw number, and every recommended action has a specific owner or team

Overall Health: 🟢 On Track / 🟡 Watch / 🔴 Action Required

Metric	Current	Target	vs. Last Period	Status
[metric]	[value]	[target]	[+/-%]	[🟢/🟡/🔴]

Dimension	0	5	10
Target & trend discipline	Metrics shown as bare snapshots; no targets or period-over-period comparison	Most metrics have targets and trends, but some RAG statuses don't follow from the numbers or targets are accepted uncritically	Every metric has target, trend, and a status that follows from both — and at least one target is itself challenged if it's no longer meaningful
Root cause depth	Movements listed without explanation ("activation dropped 27pts")	Flagged metrics have hypotheses, but they're generic ("onboarding friction") with no diagnostic to confirm them	Every flagged metric has a specific, falsifiable hypothesis plus a named diagnostic step, and at least one cross-metric correlation (e.g. activation → retention lag) is drawn
Segment honesty	Only blended aggregates reported; opposing segment trends invisible	Some segment cuts shown, but the headline observations still lean on averages that hide divergence	Every material aggregate is decomposed where segments diverge, and the divergence itself is surfaced as a finding, not a footnote
Verdict & actionability	No overall rating, or a rating asserted without evidence; actions missing or ownerless	Overall RAG present and roughly justified; actions exist but some lack owners, dates, or a link to a flagged metric	Overall rating argued from specific evidence (including against the good news), and every action has a named owner, a date, and traces to an investigation or observation