The Best Data & Analytics Skills for Claude Code (pandas, SQL, ML Pipelines)
Data work is where agent skills quietly shine — the difference between an agent that guesses at your dataframe and one that knows pandas idioms cold. I scored the full category on six dimensions. Here's what held up.
Data analysis is a discipline with right answers, which makes it ideal territory for skills: the patterns are known, the failure modes are known, and an agent that has internalized both is dramatically better than one winging it.
Seven skills earned their spot in the category evaluation.
1. vercel-optimize — 9.4/10
Cost and performance optimization for deployed Vercel projects.
Why: The surprise #1. It combines real observability data with concrete remediation steps — not generic "reduce your bundle size" advice, but specific findings tied to your actual deployment.
The one thing: It's Vercel-specific by design. If you're on another platform, it won't fire — and shouldn't.
2. ml-pipeline — 9.0/10
Designs and implements production-grade ML pipeline infrastructure: data versioning, feature stores, orchestration, model registries.
Why: Every implementation step ends with an assertion checkpoint. It's the rare skill that treats pipeline building as an engineering discipline rather than a notebook exercise.
The one thing: It designs for production from step one. For exploratory notebook work, it's heavier than you need.
3. pandas-pro — 9.0/10
DataFrame operations for analysis, manipulation, and transformation.
Why: It knows the idiom layer — the difference between a chained-assignment warning and the correct .loc pattern, between a merge that preserves index and one that silently resets it. Your agent stops producing code that works once and breaks on the next dataframe.
The one thing: It won't fix your data model. If your problem is how the data is structured rather than how it's manipulated, no pandas skill saves you.
4. postgres-pro — 8.9/10
PostgreSQL query optimization, replication configuration, index strategy.
Why: The query-optimization workflow is genuinely systematic: EXPLAIN analysis, index gap identification, statistics checks — in that order, not whatever the agent feels like.
The one thing: It needs read access to actually run EXPLAIN. Read-only credentials are enough, but without them it degrades to advice.
5. spark-engineer — 8.8/10
Writing Spark jobs, debugging performance issues, configuring Spark clusters.
Why: Spark performance is a dark art — shuffle tuning, partition strategy, broadcast decisions. This skill encodes the decision tree your agent would otherwise learn by burning cluster hours.
The one thing: The advice quality tracks your Spark version. Check the version assumption if you're on an older cluster.
6. ab-test-analysis — 8.0/10
A/B test results with statistical significance, sample size validation, and effect-size analysis.
Why: It refuses to declare a winner without checking sample size and multiple-comparison corrections — the two mistakes that make most A/B readouts fiction.
The one thing: It's honest to a fault. If your test is underpowered, it will tell you, and you won't like it.
7. cohort-analysis — 8.0/10
Retention curves, feature adoption cohorts, segment comparisons on engagement data.
Why: Clean methodology: cohort definition first, metric definition second, visualization last. Most agents draw the chart first and discover the definition afterward.
The one thing: Garbage in, garbage out. If your event data has inconsistent naming, the skill surfaces that problem rather than hiding it.
The pattern
Data skills succeed when they encode what experienced practitioners check before they act: the dtype check before the merge, the EXPLAIN before the index, the power analysis before the winner declaration. The category's weakest entries skip straight to output.
Full scorecards are on the data & analytics collection page.