Skip to content
velonx*SKILLS

Data Analysisv1.2.0featured

Explore, clean and analyse tabular data (CSV, Excel, JSON) with code, then report insights with charts and the exact steps taken. Use when the user shares a dataset or asks what the data shows, wants trends, comparisons, anomalies, or a chart.

Author
Velonx
License
MIT
Updated
Sep 25, 2026

Overview

A disciplined analysis loop: understand the data, clean it transparently, answer the actual question, and show the evidence. Every number in the report can be traced to code.

When to Use

  • "What does this data show?", "why did X drop in March?", "chart revenue by region"
  • Any CSV/Excel/JSON the user wants insights from.

For data in a live database, use sql-analysis.

Installation

coming soonVelonx CLI:velonx install data-analysis
  1. Get the skill folder

    git clone --depth 1 https://github.com/velonx/agent-skills.git
    The skill is agent-skills/skills/data-analysis/. Or open it on GitHub and download the files.
  2. Give SKILL.md to your agent as instructions

    Add it to the system prompt, custom instructions or rules file your agent reads. Keep the rest of the folder next to it if the skill refers to examples or scripts.
  3. Ask for the task in your own words

    The agent follows the skill when your request matches it. You can also name it: “use data-analysis”.

Agents move these settings between versions — if a path doesn’t match, check your agent’s docs.

Usage

  1. Clarify the question and what decision it supports. If the user just says "analyse this", propose 3 questions and pick the most useful.
  2. Profile the data with pandas: shape, dtypes, head(), missing values per column, duplicates, value ranges, date coverage.
  3. Clean transparently. Log every step (dropped rows, parsed dates, fixed types, outlier handling) with row counts before/after. Never silently drop data.
  4. Analyse with the simplest method that answers the question: group-bys, rates, period-over-period change, distributions. Use statistics (confidence intervals, tests) only when comparing groups and say what they mean in plain words.
  5. Chart only where it helps: one message per chart, labelled axes with units, a title that states the finding ("Refunds doubled after the March price change").
  6. Report:
    • Top 3 findings, each with the number and the chart or table behind it.
    • Caveats (data gaps, small samples, correlation ≠ causation).
    • Cleaning log.
    • The code, so the user can rerun it.

Examples

Prompt: "Here's our sales export. Why did Q2 revenue fall?"

The agent profiles 48k rows, finds 3% duplicate orders, removes them, splits revenue by region and product, and shows that the drop is almost entirely one product line in one region after a stock-out — with a chart and the query behind it.

Requirements & compatibility

Your agent needs

  • Run code (e.g. Python)
  • Read and write local files

Environment: Needs Python 3.10+ with pandas; matplotlib for charts.

Written for

GenericClaudeOpenAIGeminiCursor

Limitations

  • Large files (>1 GB) may need sampling; the agent says when it samples.
  • Findings are descriptive; causal claims need experiments or domain knowledge.

Files

Changelog

  • 1.2.0 — Cleaning log with before/after row counts.
  • 1.1.0 — Chart titles state the finding.
  • 1.0.0 — Initial release.

Contributing

This skill lives in velonx/agent-skills. Improvements go through a pull request, reviewed by maintainers.