analyze-drivers — Claude Code skill: find process bottlenecks and causes
analyze-drivers — a skill to identify what drives differences in process duration, wait times, resource utilization, and rework. Frames testable hypotheses and distinguishes competing explanations through attribute correlation.
MIT ★ 1,167 updated 2026-10-29
/plugin marketplace add microsoft/power-platform-skills
/plugin install process-intelligence@power-platform-skillsWhat analyze-drivers does
analyze-drivers uncovers why one process path takes longer than another or why rework spikes. The skill works with case-level metadata and attributes, ranks attribute influence on duration, waiting time, active time, utilization, event count, or loop counts.
The investigation follows an ordered trace: first validates metadata level and reuses cached signals, then runs correlation via get_correlation_v2 using influence formulas (DurationInfluence, WaitingTimeInfluence, ActiveTimeInfluence, CaseUtilizationInfluence, ReworkCountInfluence, and others), pulls attribute statistics via get_attribute_statistics if prevalence or means are needed, then tests competing hypotheses through compare-cohorts or focused analysis.
Critical: influence is a prevalence-weighted descriptive score, not causation, Pearson correlation, or significance. Results return as hypotheses with selection and confounding factors noted; the skill requires a specific defined outcome and matching case-level attributes from your context, and marks claims as tested versus merely observed.
What it helps with
- find attributes that drive long-running delays in order fulfillment
- identify which process variants generate the most rework and loops
- test two rival explanations and pick the one backed by data
- rank resource utilization differences by case-level characteristics
- connect metadata changes to event duration shifts
Example prompts for the agent
- Show me which attributes correlate with wait time in the approval step.
- Why is rework higher in the northeast region? Test competing hypotheses case by case.
- List the top three factors driving the jump in average duration last quarter and how to test them.
How to install analyze-drivers
1Plugin marketplace
/plugin marketplace add microsoft/power-platform-skills/plugin install process-intelligence@power-platform-skillsThis installs the whole process-intelligence plugin with all its skills
2Skills CLI (Vercel)
npx skills add microsoft/power-platform-skills --skill analyze-drivers -a claude-code3Manually: put the files in place
~/.claude/skills/analyze-drivers/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/analyze-drivers/
1Skills CLI (Vercel)
npx skills add microsoft/power-platform-skills --skill analyze-drivers -a codex2Manually: put the files in place
.agents/skills/analyze-drivers/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add microsoft/power-platform-skills --skill analyze-drivers -a cursor2Manually: put the files in place
.cursor/skills/analyze-drivers/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install microsoft/power-platform-skills plugins/process-intelligence/skills/analyze-drivers2Manually: put the files in place
.github/skills/analyze-drivers/SKILL.md1Skills CLI (Vercel)
npx skills add microsoft/power-platform-skills --skill analyze-drivers -a gemini-cli2Platform CLI
gemini skills install https://github.com/microsoft/power-platform-skills --path plugins/process-intelligence/skills/analyze-driversManually: put the files in place
.devin/skills/analyze-drivers/SKILL.mdThe legacy .windsurf/skills/ path still works
Claude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about analyze-drivers
How do I install analyze-drivers in Claude Code?
Run /plugin marketplace add microsoft/power-platform-skills, then /plugin install process-intelligence@power-platform-skills, then npx skills add microsoft/power-platform-skills --skill analyze-drivers -a claude-code. The skill will be available at ~/.claude/skills/analyze-drivers/SKILL.md.
What does analyze-drivers do?
The skill diagnoses what causes differences in process speed, wait times, resource use, and rework by ranking attribute influence. It builds testable hypotheses, suggests competing explanations, and specifies what evidence would distinguish between them.
What tools does analyze-drivers rely on?
The skill uses Read and AskUserQuestion. Internally, it calls get_correlation_v2 to rank attribute influence, get_attribute_statistics for grouped metrics, and focused comparisons via compare-cohorts to validate hypotheses against your frozen filter scope.
Source text
SKILL.md · MIT
---
name: analyze-drivers
description: Use when seeking explanatory attributes or hypotheses for differences in process duration, waiting, utilization or rework.
allowed-tools: Read, AskUserQuestion
---
# Drivers: association before explanation
Use the [shared contract](../../references/analysis-contract.md) and
[filter/unit reference](../../references/filters-and-units.md).
Reuse metadata and frozen baseline scope; choose a specific defined outcome and relevant case-level
attributes from the user's context, without a candidate quota. If none is known, ask; do not sweep.
State a testable hypothesis, possible competing explanation and what would distinguish them.
Load the [investigation method](../../references/investigation-method.md) for evidence/status tracking.
For utilization, first check timestamp imports and the zero-duration convention in the unit
reference. A cohort-mix convention is not resource efficiency or measured productive time.
## Ordered trace
1. Check A's metadata level before correlation. Reuse a cached signal instead of repeating it.
`get_correlation_v2 {S,attributeName:A,influenceFormula:I,sortOrder:"Descending",filterOptions:F}`.
| Outcome | I |
|---|---|
| Duration / waiting / active time | DurationInfluence / WaitingTimeInfluence / ActiveTimeInfluence |
| Utilization / event count | CaseUtilizationInfluence / EventCountInfluence |
| Rework / loops / self-loops | ReworkCountInfluence / LoopCountInfluence / SelfloopCountInfluence |
Use the advertised enum and exact spelling. There are **no** `itemsPerPage`, `itemsToSkip`,
`top`, `metrics` or `limit` arguments. The service ranks at most 20 groups.
2. Only if prevalence or mean evidence is needed and statistics is advertised:
`get_attribute_statistics {S,attributeName:A,filterOptions:F,itemsPerPage:5,itemsToSkip:0,metricToSortBy:"CaseFrequency",sortOrder:"Descending"}`.
Match returned values explicitly. Influence-ranked groups may differ from
frequency-ranked groups; an absent match is unknown, not zero.
3. For a remaining hypothesis, use a discriminating matched comparison via compare-cohorts,
focused statistics/cases, or another relevant candidate. Derive each test from frozen F,
never an outcome-only exploratory subset. A cached signal can go directly to its test.
Continue needed checks/pages without fixed analytical-call or candidate quotas.
If matched evidence contradicts the proposed direction, mark it **contradicted** within that
scope; do not privilege the earlier influence score or claim it proves a cause.
## Non-case branch
An event-level Resource cannot be sent to correlation. If useful and advertised, use the
statistics call above for that attribute, but describe grouped **event** duration/frequencies,
not case-duration drivers. Otherwise ask for a case-level candidate or report the capability gap.
Do not silently substitute a different causal analysis.
## Report and stop
Influence is a prevalence-weighted descriptive score, **not Pearson correlation, significance,
causal effect or confidence**. Its formula already scales by 100; do not multiply again.
Return at most five useful signals with available counts/means, units, scope and partial coverage.
Label explanations as hypotheses, note selection/confounding, and state one falsifiable evidence check.
Record tested versus merely observed claims, alternatives and missing evidence; revise or stop
when the next check cannot discriminate. No assumed outcome maturity or closure.
Stop when no discriminating evidence remains; do not paginate correlations or invent sample sizes.
© Microsoft · MIT · microsoft/power-platform-skills