analytics-skills

Analytics skills for Claude, Cursor, and other AI agents. Read web analytics like a senior analyst: diagnose traffic changes, judge channel quality, read funnels, declare typed events, and read A/B tests without the usual rookie mistakes. Ships with tool-maps for Amplitude, Clamp, GA4, Mixpanel, and PostHog; add your own under tool-maps/. The schema-authoring and experiment-reading skills are built on the open Event Schema spec.

What's in the box

Skill	What it does
analytics-profile-setup	One-time interview that captures your business context (model, primary conversion, traffic range, ICP, data stack) into a local analytics-profile.md. Every other skill reads this file so answers are calibrated to your industry and scale. Run this first.
analytics-diagnostic-method	The spine. Five-step method: load profile, frame the question, build a MECE hypothesis tree, triangulate, present with the Pyramid Principle. Covers signal vs noise and Simpson's paradox. Referenced by every other skill.
traffic-change-diagnosis	Drill path for "why did traffic change". Fingerprints for tracking regressions, bot spikes, deploy-correlated drops, campaign ramps, SEO decay, and platform changes. Measurement checks first, always.
channel-and-funnel-quality	Volume × engagement × conversion as a matrix. Vanity-traffic detection. Expected drop-off ranges per funnel step type. Mix-shift handling. Industry-specific benchmarks.
metric-context-and-benchmarks	What's a good bounce / engagement / duration / CVR / churn / LTV:CAC / activation, by model. When each metric lies. Minimum sample sizes before trusting a rate.
event-schema-author	Authors event-schema.yaml from existing track() calls. A portable, typed declaration of every product analytics event the codebase fires. The CLI generates a TypeScript type so call sites are autocompleted and type-checked at build time. Vendor-neutral; works with any analytics SDK.
experiment-result-reader	Read a running A/B test honestly. Pulls per-variant exposure and conversion counts, computes lift, applies sample-size and sequential-testing discipline, checks for mix-shift and sample-ratio mismatch, and returns a verdict with caveats instead of a false-positive. Reads the experiment from the experiments: section of event-schema.yaml when present.
bayesian-experiment-reader	Bayesian counterpart to experiment-result-reader. Beta-Binomial for CVR, Normal-Normal for continuous. Returns P(variant beats control), credible intervals, and expected loss. Ship-decision rule: ship if P(better)>95% AND expected loss < tolerance, instead of "stat sig yes/no."
sequential-monitoring	mSPRT and confidence sequences for honest peeking. Answers "can I call this early?" without α-inflation from looking at the dashboard every day. Pairs with both experiment readers.
causal-query-classifier	Pearl's three-rung hierarchy as a query classifier. Tags every analytics question as rung-1 (association), rung-2 (intervention), or rung-3 (counterfactual). Refuses to escalate rung-1 findings into rung-2 recommendations without an identification strategy.
causal-dag-builder	Emits a Mermaid causal DAG before answering "did X cause Y" on observational data. Applies the back-door criterion to pick the adjustment set. Surfaces confounders, mediators, and colliders explicitly instead of "controlling for everything."
causal-evidence-checklist	Bradford Hill's 9 viewpoints as a rubric before recommending a rollback/ship action. Refuses "high confidence" verdicts when fewer than ~5 criteria pass.
anomaly-detection-time-series	STL decomposition, Bayesian online changepoint detection, Prophet, quantile regression, SPRT, Granger causality. Use when traffic-change-diagnosis fingerprints match ambiguously, the change date is contested, or the user asks "is this real?"

Install

Claude Code (recommended)

/plugin marketplace add clamp-sh/analytics-skills
/plugin install analytics-skills@clamp-sh

Skills auto-load and become available as /analytics-skills:<name> when the task matches.

Other clients

• Cursor / Copilot / any Skills-spec client: npx skills add clamp-sh/analytics-skills
• Codex CLI: git clone https://github.com/clamp-sh/analytics-skills.git then cp -R analytics-skills/skills/* ~/.codex/skills/
• claude.ai: download per-skill zips from the latest release and upload via Settings → Features → Skills
• Manual (any Claude Code, no plugin): git clone then cp -R skills/* ~/.claude/skills/

Usage

These are model-invoked skills. You don't need to call them by name; just ask real analytics questions and Claude will load the relevant skill.

First run, on a new project:

Set me up. I want the skills calibrated to my business.

This loads analytics-profile-setup, walks a 5-minute interview, and writes analytics-profile.md to the repo root. Every subsequent question gets industry-aware answers.

After setup, ask real questions:

Traffic dropped 30% on Tuesday. What happened?

Loads traffic-change-diagnosis. Walks the hypothesis tree (measurement → time-shape → channel → cohort → content), pulls numbers from your analytics source, returns a diagnosis. Not a screenshot of a chart.

Is our signup funnel broken? Pricing → checkout is converting at 14%.

Loads channel-and-funnel-quality and metric-context-and-benchmarks. Compares against expected step drop-off, slices by cohort, flags whether 14% is low, normal, or suspicious given your sample size and model.

Our bounce rate is 68%. Is that bad?

Loads metric-context-and-benchmarks. Handles the GA4 vs UA definition gotcha, looks up the relevant page-type range, flags the sample-size caveat if needed.

Supported analytics platforms

The skills are platform-neutral; per-platform MCP invocations live in tool-maps/. One file per supported analytics tool, all covering the same canonical 17-row workflow taxonomy.

Tool	Tool-map	Surface
Amplitude	amplitude.md	query_amplitude_data covers most rows; cohorts and experiments are dedicated tools
Clamp	clamp.md	dedicated MCP tool per row of the canonical taxonomy
GA4	ga4.md	run_report for aggregate rows; funnels and cohort retention not exposed by the wrapper
Mixpanel	mixpanel.md	Run-Query types (insights, funnels, flows, retention)
PostHog	posthog.md	Trends/Funnels/Retention/Paths insights plus HogQL

The full row-by-row coverage matrix is at tool-maps/capability-matrix.md. analytics-profile-setup records the active platform in analytics-profile.md under tool_map:; downstream skills load the matching tool-map automatically.

Using a different analytics source? The method still applies. Add a tool-map for it (see tool-maps/README.md for the template).

Contributing

Issues and PRs welcome. See CONTRIBUTING.md for how to add a skill and the style conventions we follow.

License

MIT. See LICENSE.

References

Frameworks and benchmarks the skills lean on:

• Method: Minto's MECE and Pyramid Principle (1985), Simpson (1951) for the mix-shift trap, standard 95% CI sample-size rules for noise-vs-signal calls.
• Benchmarks: Unbounce 2024 (57M conversions), Wordstream 2025, Ruler 2025 (100M+ data points by industry), Littledata Shopify 2023, Imperva Bad Bot Report, Mixpanel, ChartMogul, David Skok SaaS Metrics 2.0.
• Definitions: GA4 metric definitions taken from Google's docs (note: GA4 bounce rate is not the UA single-pageview bounce rate).