Paste your metrics table. Your browser scores every metric as you type - change, target,
trend, streak and how unusual the latest value is - free, before you sign in. Then the
review reads the scorecard, and the drill-down splits the
metric that moved by segment to show whether the segments changed or the mix did.
Each example has a saved model run for both the review and the drill-down, so you can see the whole page for free.
Reviewing
Analysis is AI-generated - please confirm before acting
metrics read
anomalies
actions
prescan flags answered
disagreements
Scorecard, read
Ref
Metric
Latest
Change
Browser
Review
Reading
Bright spots
Areas of concern
Anomalies
Recommended actions
Type
Action
Metric
Success looks like
Context and caveats
Prescan flags, answered
Reconciliation with your table
Every metric had to be read once, each status held to the browser's (green against red is a disagreement), every metric two standard deviations from baseline explained, every ref real, every prescan flag answered, and every figure in the prose looked up in what was sent.
Analysis is AI-generated - please confirm before acting
top-line change
rate / mix
drivers
prescan flags answered
disagreements
Drivers
Ref
Segment
Stated
Browser
Why
Offsets (moved against the line)
Hypotheses to test
Validation
Next queries
Recommendation
Prescan flags, answered
Reconciliation with the breakdown
The drill had to be about the metric sent, every driver's contribution had to be the browser's number with the largest first, offsets had to move against the line, the verdict had to follow the rate/mix split, every prescan flag answered, and every figure in the prose looked up in what was sent.
Raw model reply
The reply did not parse as the structured result, so it is shown as it arrived.
Your recent runs
If that did not work
Check that the metrics table has a name column and at least two period columns, that you are
signed in, and that your balance covers the amount reserved next to the button. The three
examples always work and cost nothing, so they are the quickest way to tell whether the
problem is your table or the service.
What this does, and what it does not
The prescan reads your table and does the arithmetic a metrics review rests on: the latest
value against the one before (in points for a percentage), attainment against target, a status
that respects whether higher or lower is better, a least-squares trend over the last eight
periods, the current streak, and a z-score of the latest value against its own earlier
periods. It also notes which lifecycle stages - acquisition, activation, engagement, retention,
monetization, satisfaction - the scorecard has no metric for.
The drill-down reads a segment breakdown. For a rate over a base, the change splits exactly into
a rate effect (segments converting differently), a mix effect (the blend of segments shifting,
measured against the prior average) and an interaction; for a volume, each segment's
contribution. It checks that the segments add back to your scorecard before anyone trusts them.
It knows nothing you did not paste: no analytics connection, no benchmarks. Causes the model
offers are hypotheses with tests, not findings. Derived from the agent skills
@anthropics/metrics-review
and @anthropics/analyze
(anthropics/knowledge-work-plugins, Apache-2.0). The example products are fictional.