English | 中文
This is an Agent Skill for data analysis — it runs across multiple AI coding tools instead of being tied to any single one. Hand it a data file (CSV / Excel / JSON / TSV), say "analyze this", and it works like a data analyst who knows your industry: it first confirms the business context with you, then cleans and analyzes the data, and finally delivers a single-file interactive HTML report — not isolated charts and numbers, but a data story with insights and conclusions you can take straight into a meeting.
Works in: Claude Code (primary, fully tested) · Codex CLI (adapted) · and any AI coding tool that loads the same Agent Skill structure (SKILL.md + frontmatter + subdirectory resources). The five-phase workflow and statistical discipline are identical across tools — only the task-list and questioning mechanics adapt per tool.
See it first: open skills/data-analysis-workspace/sample_report.html in your browser (internet access required — chart libraries load from CDN).
- "Here's our Q2 order data — give me an analysis report" → a complete narrative report organized from a retail perspective: GMV, average order value, repeat purchases, returns — every chart comes with an interpretation, conclusions first
- "Check whether variant B in my A/B test is actually better" → a rigorous statistical answer (p-value, confidence interval, assumption checks) without forcing you through the full report pipeline
- "Analyze my account's posts from the past three months" → interpreted through a content-creator lens: engagement rate, completion rate, follower growth efficiency
- "Do a business review of this budget execution sheet" → analyzed from an FP&A perspective: revenue, gross margin, expense ratios, budget attainment
Prerequisite: at least one supported AI coding tool installed — Claude Code and/or the Codex CLI. A single command installs into whichever ones it detects (Codex only when ~/.codex/ exists).
# Run from the repo root; the skill is installed to ~/.claude/skills/data-analysis-skill,
# and to ~/.codex/skills/data-analysis-skill when ~/.codex/ is detected
bash sync-to-user.sh
# If your skills directories live elsewhere:
DATA_ANALYSIS_CLAUDE_SKILL=/your/path bash sync-to-user.sh
DATA_ANALYSIS_CODEX_SKILL=/your/codex/path bash sync-to-user.shUnder Codex, task tracking uses update_plan and the stage-1 industry confirmation appears as a numbered text question (Codex has no structured questioning tool); the workflow and statistical discipline are otherwise identical.
To upgrade later, pull the latest code and run the same command again.
-
Provide data, state your need. In your AI coding tool (Claude Code, Codex, …), point to your data file and say what you want:
Analyze this sales data for me: retail_orders.csv — I need a report -
Answer a few multiple-choice questions. Before analyzing, it confirms your industry, analysis goal, and report audience. Don't skip this — it determines what the report analyzes and who it speaks to: an executive summary for your boss, or a detailed diagnosis for the ops team.
-
Get the report. It walks through data understanding → cleaning → metric computation → analysis → report generation, and delivers an HTML file you open in a browser.
Just have a quick question? Things like "compute the correlation between these two columns" or "is group B's conversion significantly higher" won't be dragged through the full pipeline — it computes and answers directly, with the same statistical rigor.
- A single HTML file: double-click to open, interactive charts (zoom, tooltips), ready to share with colleagues (they need internet access to open it too)
- A fixed six-section structure, conclusions first: Key findings (3–5) → Background & goals → Data overview → Deep-dive analysis (sectioned by industry themes) → Conclusions & recommendations → Appendix: data processing notes
- Every chart comes with an interpretation: what the chart shows and what it implies — it doesn't just dump charts on you
- Traceable processing: how missing values were handled, whether outliers were dropped — every decision is recorded in the appendix
- Report language follows your prompt language: ask in English, get an English report
| Industry | Typical scenarios | What the report focuses on |
|---|---|---|
| Retail / e-commerce | Marketplaces, online stores, offline/omnichannel retail | GMV, average order value, repeat purchase rate, return rate, category mix |
| Internet / SaaS | Subscription software, internet products, membership services | MRR/ARR, retention, churn, net revenue retention |
| Content creators | Account operations on Xiaohongshu (RED), Douyin, Bilibili, WeChat official accounts, etc. | Engagement rate, completion rate, net follower growth, follower conversion efficiency |
| Corporate finance / FP&A | P&L, expenses, budget, cash flow analysis | Revenue growth, gross margin, expense ratios, budget attainment |
| Funds | Fund product performance and operations data | NAV returns, max drawdown, subscriptions/redemptions, holder structure |
| Real estate | Property developer sales and market data | Sales & absorption, cash collection, market volume-price trends |
| Construction | Construction/contracting company operations | Project output & progress, labor/material/machinery costs, contract intake & collections |
| Education & training | Training institutions, online education platforms | Enrollment & renewals, attendance & completion, channel acquisition & CAC |
| Everything else | When none of the above fits, it falls back to a general lens | Scale, trends, composition, distributions — without force-fitting another industry's KPIs |
When the task calls for specialized methods — A/B testing, time-series forecasting, causal inference, machine learning — it loads the corresponding method module on demand; you don't have to ask.
The skill works under professional discipline — it would rather say less than make things up:
- No fabricated industry benchmarks. Unless you provide benchmark data or explicitly authorize a web search, it only compares within your own data — it never quotes an industry average "from memory."
- Correlation ≠ causation. Without causal-method validation it won't claim "A drove B," and it proactively calls out potential confounders.
- No forced conclusions from small samples. When the sample is too small, it tells you plainly that confidence is limited — even if "the boss needs an answer today."
- Bad data gets a diagnosis first. If data quality is too poor, it tells you what's missing and how to fix it, instead of fabricating a polished-looking analysis.
- No live dashboards / dashboard apps (the deliverable is a static single-file report)
- No web scraping — it won't fetch data for you
- No direct database connections — export your data to CSV / Excel files first
The report opens blank / charts don't render? Chart libraries load from CDN, so opening the report requires internet access. The CDNs are chosen to be reachable from mainland China.
Want a PDF or static images instead? Just say so ("give me static charts" / "export to PDF") and it switches to a static-chart approach.
Do I need to set up Python or install packages myself? Analysis runs locally inside your AI coding tool using Python. If dependencies like pandas are missing, it tries to install them automatically, with fallbacks if installation fails — generally nothing for you to worry about.
It asks about my industry and goals — can I skip that? The confirmation step before a full report is deliberate: data that looks like retail data doesn't mean you want a retail reading of it. Ten seconds of choices make the report fit much better. Quick single questions skip this step entirely.
Released under the MIT License © 2026 cabbage2000-lab.