Works with the AI Agents you already use
Claude Code
Cursor
Codex CLI
Muse
Dots+22 moreClaude CodeCursorCodex CLIMuseDots
OpenClaw
Manus
Grok
Grok Bot+18 agentsMetric Drop Investigator
Investigate why a metric fell: check tracking, compare segments, test likely causes, and choose the next investigation.
$14.40
Metric Drop Investigator
Example session with this skill installed
SignalCart: a conversion-rate investigation
All companies, observations, dates, customer quotes, and numerical values in this fixture are fictional simulation data. They do not describe a real company or a measured customer result.
Source packet
Read fixtures/metric-windows.csv. Each row is a disjoint platform-period aggregate. Eligible sessions are deduplicated sessions eligible to check out. Completed means an authoritative completed order within the same seven-day session cohort. raw_completed and transformed_completed are count reconciliation fields for the same population. Gross orders are used; refunds are outside this exercise.
Before window: 2026-09-01 through 2026-09-07. After window: 2026-09-08 through 2026-09-14. A UI fix was deployed on 2026-09-15, after both windows. No traffic-campaign log, randomized experiment, or user replay has been supplied. The dashboard reports a decline; the operator wants to know where to investigate first.
Buyer request
Explain the overall and within-platform movement, reconcile raw/transformed counts, separate mix from conversion changes, and write a hypothesis board with confidence and one discriminating next test. Use only the supplied synthetic data. If running the optional helper, give the CSV path explicitly; it uses only the supplied local file and prints calculations.
Additional practice
fixtures/within-segment-decline.csv is a second synthetic cohort with the same metric definitions and comparison windows. Investigate it independently rather than reusing the first diagnosis.
- Read your context and instructions
- Compiled the metric drop investigator
- Generated the data export
SignalCart metric investigation — synthetic demonstration
The conversion rate fell from 40% to 28%: a 12 percentage-point decline, or 30% relative decline. In this supplied cohort, platform composition explains the entire aggregate movement: mobile's share increased from 50% to 80%, while mobile conversion stayed at 20% and web conversion stayed at 60%. This is a high-confidence arithmetic finding. The cause of the composition change is unknown.
Decision: investigate entry/source composition and eligibility before prioritizing a conversion-path repair. Proposed owners are the analytics lead for evidence gathering and product lead for prioritization; no actual owner was supplied. The September 15 UI fix happened after both comparison windows and cannot explain the start of this decline. Its effect is unmeasured here.
All data is fictional. This analysis uses only packages/metric-drop-investigator/fixtures/source-packet.md and fixtures/metric-windows.csv. No live system was accessed and no changes were made to a product, dashboard, or account.
Metric contract
| Field | Contract |
|---|---|
| Behavior and business relevance | Eligible sessions create completed orders. The decline represents 240 fewer gross completed orders at the same overall eligible-session count; revenue and net order value are unavailable. |
| Numerator | Authoritative completed orders within the supplied seven-day session cohort, as represented by completed. |
| Denominator | Deduplicated eligible checkout sessions, as represented by eligible. |
| Unit and population | Eligible session; four disjoint platform-period aggregates. Mobile and web cover the entire supplied fixture, with no additional eligibility rules supplied. |
| Comparison | Before: September 1–7, 2026. After: September 8–14, 2026. Both windows span seven days. |
| Source | Synthetic CSV aggregates; raw_completed and transformed_completed reconcile the same cohort's completed count. |
| Join and deduplication | Sum the disjoint platform cells by period. Session deduplication is a source-packet assertion; no underlying session IDs, retry records, or correction history are available to audit it. |
| Gross/net | Gross completed orders; refunds are outside this exercise. |
| Coverage and freshness | Four rows, two platforms, two weekly cohorts; all supplied rows used. Extraction time, reporting lag, daily observations, and late-arrival history are unavailable. |
| Blind spots | No campaign/source ledger, raw event records, adjacent funnel events, replay, randomized experiment, geographic cuts, release history before the decline, or downstream durability data. |
Observed movement
These are instrumented facts within the synthetic fixture, rather than measurements of a real company.
| Population | Before eligible | Before completed | Before rate | After eligible | After completed | After rate | Rate change | Relative rate change |
|---|---|---|---|---|---|---|---|---|
| Mobile | 1,000 | 200 | 20% | 1,600 | 320 | 20% | 0 pp | 0% |
| Web | 1,000 | 600 | 60% | 400 | 240 | 60% | 0 pp | 0% |
| Overall | 2,000 | 800 | 40% | 2,000 | 560 | 28% | −12 pp | −30% |
Overall eligible volume is unchanged. Mobile eligible sessions and completed orders each rose 60%; web eligible sessions and completed orders each fell 60%. Overall completed orders fell by 240, or 30%. The aggregate rate is a ratio of summed counts, not an unweighted mean of platform rates.
| Platform | Before eligible share | After eligible share | Share change |
|---|---|---|---|
| Mobile | 50% | 80% | +30 pp |
| Web | 50% | 20% | −30 pp |
Using before-period platform rates, the after-period composition predicts 0.80 × 20% + 0.20 × 60% = 28%. Using the before-period composition with after-period platform rates produces 0.50 × 20% + 0.50 × 60% = 40%. The exact decomposition is:
- Composition contribution:
(0.80 − 0.50) × 20% + (0.20 − 0.50) × 60% = −12 pp. - Within-platform rate contribution:
0.80 × (20% − 20%) + 0.20 × (60% − 60%) = 0 pp.
This decomposition establishes where the observed arithmetic change lies. It does not establish what caused the mix change or exclude offsetting subsegment changes inside each platform.
Count reconciliation and timing
| Period/platform | Completed | Raw completed | Transformed completed | Difference |
|---|---|---|---|---|
| Before/mobile | 200 | 200 | 200 | 0 |
| Before/web | 600 | 600 | 600 | 0 |
| After/mobile | 320 | 320 | 320 | 0 |
| After/web | 240 | 240 | 240 | 0 |
All four same-cohort cells reconcile, including period totals of 800 and 560. This rules down loss of completed counts between the supplied raw and transformed fields. It does not audit eligibility instrumentation, event lineage, or unseen source errors.
The movement is visible in the September 8–14 aggregate relative to September 1–7. These weekly rows cannot locate the first affected day, distinguish a sudden change from a gradual one, or establish statistical significance across independent sessions. September 15 is the only supplied release date and lies after both windows. There is no post-fix cohort with which to assess recovery.
Hypothesis board
| Hypothesis | Position | Evidence for | Evidence against / unknowns | Rule-in / rule-out test | Verdict and confidence |
|---|---|---|---|---|---|
| Platform mix explains the aggregate decline | Entry / routing / eligibility | Mobile share rises 30 pp; the exact weighted-rate change is −12 pp. | Cause of the share movement is unobserved. | Reweight using a fixed platform mix; a residual rate decline would rule down a complete composition explanation. | Confirmed arithmetic explanation for these cells; high confidence. Fixed-mix rate is unchanged. |
| A source or eligibility change caused the mix shift | Entry / eligibility | Eligible platform counts redistribute while the overall count is stable. | No source, campaign, routing, or eligibility history. | Reconcile source-by-platform eligible sessions under a fixed eligibility definition in both windows; unchanged authoritative composition would rule down a real entry mix change. | Plausible mechanism, low causal confidence. |
| Conversion-path deterioration lowered platform rates | Routing / render / outcome | An aggregate decline exists. | Both observed platform rates are unchanged; smaller subsegments are unavailable. | Compare fixed source/cohort cells and adjacent funnel rates; systematic within-cell deterioration would rule it in. | Ruled down at platform aggregate level; finer effects remain unknown. |
| Transformation loss caused the completion decline | Pipeline | A reported decline can arise from processing failures in principle. | Every raw, transformed, and authoritative completed count agrees. | Audit same-cohort event lineage; a raw-to-transformed shortfall would rule it in. | Ruled down for the supplied completion counts, high confidence within this coverage. |
| September 15 UI fix explains the decline's start | Render | A UI change date is supplied. | It occurs after both windows; no earlier related change is supplied. | A deployment before the affected cohort would be required. | Ruled out as the initiating cause on the supplied timeline. Fix efficacy remains unknown. |
Completed tests and one next test
Completed: weighted overall rates, eligible and completed volumes, platform rates, composition shares, exact mix/within-platform decomposition, same-cohort raw/transformed reconciliation, and release ordering. The optional packaged helper ran successfully; its output was independently checked using exact rational arithmetic.
Recommendation — one discriminating next test: the analytics lead should obtain a read-only source-by-platform eligible-session ledger for the same two windows, apply one unchanged eligibility rule, and reproduce the platform weights and source-standardized conversion rates. Complete this before the next prioritization decision; a specific calendar deadline requires an actual owner.
If the ledger reproduces the move toward mobile and source-standardized rates remain stable, focus investigation on the entry/routing change that redistributed sessions. If the authoritative ledger does not reproduce the platform shift, investigate eligibility or reporting definitions. If rates decline within stable source/platform cells, the coarse platform mix result is insufficient and conversion-path investigation becomes the priority. No campaign or code cause is asserted in advance of this test.
Reproduction
From the package directory, with Python 3 available:
python scripts/calculate_rates.py fixtures/metric-windows.csv
The helper uses only the explicitly named local CSV and Python's standard library. It calculates rates and count agreement; the decomposition, causal limits, and next-test decision above require analysis beyond its printed output.
metric-windows.csv
CSV · data export
Example file from a real run - the skill writes it into your workspace.
Connects securely to your tools. The creator never sees your data.
About this skill
What it helps you do
Investigate why a metric fell: check tracking, compare segments, test likely causes, and choose the next investigation.
Built for product managers, product leaders, founders, and analytics practitioners who need a concrete work product they can inspect and revise.
What is included
- A focused SKILL.md operating workflow with evidence and decision boundaries.
- Original fictional practice material with explicit source definitions and constraints.
- A delivery review for the errors that can change the decision.
- Linked templates or output contracts and a transfer exercise for changed evidence.
- Synthetic CSV fixtures and an optional standard-library Python calculation helper.
How to use it
Give your agent the decision, source material, audience, and requested format. Ask it to use this skill. For a first run, use the included fictional practice packet. The agent separates supplied facts from assumptions, produces the deliverable, and checks decision-critical gaps.
Why this edition
This paid edition extends the creator's original workflow with task-specific practice material and delivery resources. It is intended to help you produce and review the work, not just fill headings. The listing demo uses fictional evidence; it is not an actual customer result.
Boundaries
Works from supplied or authorized evidence; it cannot invent access, facts, benchmarks, or measured business impact. Synthetic fixtures demonstrate the method only. Quality depends on source coverage and the executing agent. Real document page limits require a supported renderer. No customer accounts are required for supplied-context practice. Optional connected research and external writes depend on your tools and authorization.
How to use
Skills on Agensi run inside your AI agent. Copy the prompt, paste it into your agent, and it connects to Agensi and installs this skill for you. If a payment is needed, your agent shows it to you first.
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
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scanned
Verified clean 6 days ago
- Passed all security checks, Safe to install