- Home
- Skills
- Data & Databases
- KPI Definition Drift Triage
Works with the AI tools you already use
KPI Definition Drift Triage
Compare conflicting KPI definitions, identify semantic drift, and produce a reconciliation plan plus a draft metric contract.
$7
KPI Definition Drift Triage
Example session with this skill installed
Our synthetic August executive dashboard reports a 4.8% "conversion rate," while the Finance workbook reports 4.1%. The dashboard uses paid orders divided by web sessions, UTC calendar-month boundaries, and excludes test traffic. Finance uses settled orders net of same-month refunds divided by eligible users, America/New_York boundaries, and a month-end snapshot refreshed one day later. Both owners say their number is correct. Is the data pipeline broken, and what must we align before publishing the board report?
- Read your context and instructions
- Compiled the kpi definition drift
- Generated the document
KPI Definition Drift Report
STATE: DEFINITION_DRIFT_CONFIRMED
Confidence: High that the documented contracts differ; no conclusion yet about pipeline health.
The two values are not calculations of one shared contract. Material differences in numerator, denominator, time boundary, refund treatment, entity grain, and refresh cutoff can explain part or all of the gap. The supplied evidence does not establish that either implementation is defective.
Value snapshot
| Source | Displayed value | Period | Unit | Refresh evidence |
|---|---|---|---|---|
| Executive dashboard | 4.8% | Synthetic August | Percent | Dashboard snapshot supplied; exact timestamp unknown |
| Finance workbook | 4.1% | Synthetic August | Percent | Month-end snapshot refreshed one day later |
Contract diff
| Dimension | Executive dashboard | Finance workbook | Result | Likely impact |
|---|---|---|---|---|
| Numerator | Paid orders | Settled orders net of same-month refunds | DRIFT | Captured, settled, and refunded orders need not match |
| Denominator | Web sessions | Eligible users | DRIFT | The ratios measure different populations and grains |
| Time basis | UTC calendar month | America/New_York calendar month | DRIFT | Boundary events can move between periods |
| Refund treatment | Not supplied | Same-month refunds removed | UNKNOWN / DRIFT | Dashboard treatment must be verified |
| Grain | Session-to-order | User-to-order | DRIFT | Repeat sessions and repeat orders affect the rates differently |
| Refresh cutoff | Exact timestamp unknown | Month-end plus one day | UNKNOWN | Late arrivals may add a secondary freshness difference |
| Test traffic | Excluded | Not supplied | UNKNOWN | Finance eligibility rules must be checked |
Ranked causes
- Confirmed: Different denominators mean the two metrics are not semantically equivalent.
- Confirmed: Paid versus settled-net-of-refund numerator logic changes inclusion.
- Confirmed: UTC versus America/New_York changes period boundaries.
- Possible: Refresh cutoff and late-arriving records add a secondary data-timing difference.
- Possible: Test-traffic treatment differs because the Finance rule is unknown.
Minimum verification checks
- Recalculate both outputs using one frozen event set and one shared period boundary.
- Produce counts for paid orders, settled orders, refunded orders, sessions, and eligible users before calculating either ratio.
- Compare distinct users, sessions, and orders to expose grain and repeat-activity effects.
- Align test-traffic filters and record the exact refresh cutoff for both sources.
- Re-run each existing formula unchanged after the aligned snapshot to separate semantic drift from pipeline defects.
Draft canonical contract — decision proposal, not approved truth
- Proposed name: Monthly purchaser conversion rate
- Purpose: Measure the share of eligible users who complete at least one settled paid order in the reporting month
- Draft numerator: Distinct eligible users with at least one settled paid order
- Draft denominator: Distinct eligible users with a qualifying visit in the same period
- Time basis: One business-approved timezone and explicit inclusive/exclusive boundaries
- Refund rule: Requires owner decision on same-month versus lifetime attribution
- Grain: Distinct user
- Owner: Unresolved; executive reporting and Finance owners must designate one accountable approver
Unresolved decisions
- Is the intended business concept session conversion, purchaser conversion, or order frequency?
- Which status establishes completion: paid, settled, or another accounting state?
- Which timezone and refund-attribution policy governs board reporting?
- Which team owns the canonical contract and version history?
Reconciliation readback
After the owners approve one contract, calculate both sources from the same frozen population, time boundary, status rule, refund rule, and refresh cutoff. The figures should then agree within an explicitly approved rounding tolerance. Until those decisions are made, do not label either existing value as the canonical KPI.
kpi-definition-drift-triage.pdf
PDF · document
Example file from a real run - the skill writes it into your workspace.
Connects securely to your tools. The creator never sees your data.
What you get
About this skill
The problem
Two dashboards can use the same KPI name and still calculate different business concepts. Changes in population, denominator, time basis, status, refunds, currency, grain, deduplication, or refresh rules can create a misleading disagreement—or hide semantic drift behind matching numbers.
What it does
- Normalizes two or more KPI definitions into comparable metric contracts.
- Labels each material dimension as MATCH, DRIFT, or UNKNOWN and cites the supplied evidence.
- Classifies the case as MATCHED_DEFINITION, DEFINITION_DRIFT_CONFIRMED, DATA_OR_PIPELINE_MISMATCH_LIKELY, INSUFFICIENT_EVIDENCE, or NOT_COMPARABLE.
- Separates semantic differences from refresh lag, data quality, join, deduplication, and transformation issues.
- Produces minimum verification checks and a draft canonical metric contract for owner review.
What you provide
Redacted KPI names and values, formulas, numerator and denominator, reporting period, timezone, population, filters, status rules, currency and refund treatment, grain, deduplication logic, source and refresh timestamps, plus small redacted query fragments when needed. Never provide credentials, unrestricted warehouse access, full customer exports, or personal data.
What you receive
- A primary state and confidence assessment.
- A value snapshot and dimension-by-dimension contract diff.
- Ranked confirmed and possible causes.
- The minimum checks needed to discriminate definition drift from data or pipeline problems.
- A draft canonical contract, unresolved policy decisions, and an exact reconciliation readback.
Why this is different
The skill does not declare the newer, larger, or executive-facing number correct. It also does not treat matching values as proof that the underlying definitions match. Missing evidence remains UNKNOWN, and unresolved business policy is returned for an accountable owner to decide.
Known limitations
It uses only the redacted evidence the user supplies. It does not access warehouses, dashboards, files, APIs, credentials, or production systems, and it does not rewrite queries or reports. A draft canonical contract is a decision proposal, not approved organizational truth.
How to install
Works the same in every agent - Claude, Cursor, Codex, Copilot and 20+ more.
- 1
Download the ZIP
Free skills download straight away. Paid skills unlock right after purchase.
- 2
Unzip into your skills folder
Every agent reads skills from one folder on your machine. Drop the unzipped folder in there.
- 3
Ask your agent to use it
Restart the agent if it was already running. It picks the skill up automatically - no config needed.
Skills folder by agent
Click the path to copy it. Create the folder if it does not exist yet.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill passes our trust checks below.
Security scanned
Passed our 8-point scan before listing
Fresh listing
Recently published to Agensi
30-day refund
Not a fit? Get your money back
Trust & safety
Security scanned
Verified clean 17 days ago
- Passed all security checks, Safe to install