Docs & Knowledge · coffeefuelbump/csv-data-summarizer-claude-skill

csv-data-summarizer

Name: csv-data-summarizer
Author: coffeefuelbump/csv-data-summarizer-claude-skill

Analyzes CSV files, generates summary stats, and plots quick visualizations using Python and pandas.

L4 High-Privilege · Graded auth & full auditV1 Commit-PinnedExecute CommandInstall DepsWrite FileDual-pass Audit

Author coffeefuelbumpGitHub coffeefuelbump/csv-data-summarizer-claude-skill ↗Stars 465License 仓库内无 LICENSE 文件（git ls-tree 无 LICENSE）；README 自述 'MIT License - feel free to use this skill for personal or commercial projects!'，GitHub API license=nullCommit 9b3affd270

Control levels describe skill capabilities and suggested control measures; they do not represent quality or safety certifications, nor do they dictate usability. Usage decisions rest entirely with users and administrators. Different agent hosts apply varying constraints; do not assume hosts provide automatic protection. Third-party skill copyrights belong to their original authors. Use at your own risk.

1How It Works & Mechanism

The entire capability is implemented in a single 182-line pandas function `summarize_csv(file_path)`: read CSV → route through fixed branches by column type → assemble a string report, while saving plots via `plt.savefig` directly into the **current working directory** during execution.

analyze.py

def summarize_csv(file_path): """ Comprehensively analyzes a CSV file and generates multiple visualizations. Args: file_path (str): Path to the CSV file Returns: str: Formatted comprehensive analysis of the dataset """ df = pd.read_csv(file_path)

The report is built as a string array with piecewise appends (dozens of `summary.append(...)`), ending with `return "\n".join(summary)`. Plot file names and report text are hardcoded constants (correlation_heatmap.png / time_series_analysis.png / distributions.png / categorical_distributions.png).

Branching logic is strictly driven by column types or column name substrings rather than business or data semantics: numeric columns trigger statistics and correlations, object columns trigger value frequencies, and columns with 'date' or 'time' in their names trigger time-series analysis.

analyze.py

date_cols = [c for c in df.columns if 'date' in c.lower() or 'time' in c.lower()]

Supporting code: `numeric_cols = df.select_dtypes(include='number').columns.tolist()`, `categorical_cols = df.select_dtypes(include=['object']).columns.tolist()`, and columns containing 'id' in their names are subsequently filtered out (`categorical_cols = [c for c in categorical_cols if 'id' not in c.lower()]`).

The report content and sequence are fixed: dataset overview (shape/column names) → per-column dtype → missing value statistics → numerical describe → correlation matrix → categorical Top 10 values → time-series range → generated chart list.

analyze.py

summary.append("📊 DATA OVERVIEW")

The missing values section reports counts and percentages per column (`Missing values: {missing:,} ({missing_pct:.2f}% of total data)`), outputting `✓ No missing values - dataset is complete!` when no missing values exist.

Visualizations consist of four conditional charts: correlation heatmap (≥2 numeric columns), time-series line chart (date column present and numeric columns present, up to 3 series), numeric distribution histograms (2x2 grid, up to 4 columns), and categorical horizontal bar charts (2x2 grid, up to 4 columns).

analyze.py

plt.savefig('correlation_heatmap.png', dpi=150) plt.close() charts_created.append('correlation_heatmap.png')

All charts use `plt.savefig(...)` to relative paths followed by `plt.close()`. Maximum count limits are hardcoded via slices (`categorical_cols[:5]` in text, `value_counts.head(10)`, `numeric_cols[:3]/[:4]`).

'Do not ask the user' is mandated as a top-priority strict behavioral rule: SKILL.md uses all-caps paragraphs forbidding all questions and options, demanding immediate execution of the full analysis and delivery in a single turn.

SKILL.md

**DO NOT ASK THE USER WHAT THEY WANT TO DO WITH THE DATA.** **DO NOT OFFER OPTIONS OR CHOICES.**

The same section lists prohibited phrases ('What would you like me to do with this data?', 'Here are some common options:', etc.) and prohibited behaviors (listing options, awaiting user instructions, providing partial analysis). This represents the skill's clearest product philosophy: eliminating analysis-assistant confirmation fatigue through 'zero interaction'.

Trigger contexts are explicitly specified (uploading/referencing a CSV, asking to summarize/analyze/visualize, wanting to inspect data structure and quality), along with rules for detecting missing values and date columns.

SKILL.md

Claude should use this Skill whenever the user: - Uploads or references a CSV file - Asks to summarize, analyze, or visualize tabular data

Invocation is documented as 'The Skill provides a Python function summarize_csv(file_path)'—meaning the host Claude writes and executes code calling this function, rather than relying on a built-in CLI subcommand (analyze.py only accepts an optional path argument in `__main__`, defaulting to 'resources/sample.csv').

Distribution is handled via a zip archive uploaded to Claude.ai Skills, while identical source code is maintained in the repository; files inside the zip match the repository workspace byte-for-byte (verified).

README.md

2. Go to [Claude.ai](https://claude.ai) → Settings → Capabilities → Skills 3. Upload the zip file 4. Enable the skill

Verification: SHA-256 hashes of analyze.py, SKILL.md, and requirements.txt inside the zip match identical repository files exactly (see verification.second_pass). The zip contains SKILL.md, analyze.py, requirements.txt, and resources/{sample.csv,README.md}, mirroring the workspace structure minus examples/.

2Core Capabilities

01Generates a complete data health-check report in a single invocation: row/column dimensions, per-column types, missing value distribution, numeric descriptive statistics, and correlation matrix

02Automatically renders conditional visualizations (correlation heatmap / time-series line plot / numeric distribution histograms / categorical Top-N bar charts)

03Categorical column frequency breakdown (first 5 categorical columns, Top 10 values and percentages per column)

04Automatic date column identification and interval summary (identified when column names contain 'date' or 'time')

05Column-level missing value breakdown (counts + percentages) as part of data quality evaluation

06Runs independently outside the host environment (analyze.py includes a `__main__` entrypoint, defaulting to resources/sample.csv)

07Zero-interaction execution discipline: forbidden from asking questions, forbidden from offering options, required to return full results immediately

3External Dependencies

Kind	Dependency / Endpoint
package	pandas
package	matplotlib
package	seaborn
network	无：脚本不发起任何网络请求（无 URL 取数、无遥测、无外部 API）；仓库内 URL 仅出现在 README/resources 的徽章与社区链接中

4Recommended Governance & Evidence

Suggested Controls for Safe Operation

For Agent Runtime

Graded/per-use authorization, sandbox recommended

For Administrators

Approval workflow, rollback support, full audit

Evidence & Triggers (File Paths & Excerpts)

scripts/: Executes local script or command: python analyze.py [csv_path] (defaults to resources/sample.csv)
scripts/ 或 SKILL.md: Declares dynamic installation of external language package dependencies
本地工作目录: File I/O and artifact output: writes PNG files with fixed filenames to the **current working directory**: correlation_heatmap.png, time_series_analysis.png, dis

Audit Notes: Based on unified tiering: this skill is classified under 'Local-only tool / local file operations'—it utilizes pandas/matplotlib/seaborn locally for statistics and plotting, with no network outbound traffic, credential access, subprocessing, or arbitrary code execution surface; dependencies are official PyPI packages. It does not qualify for green tier solely because it executes Python scripts, writes fixed-name image files to the working directory (risking file overwrites), and requires installing three third-party libraries. With zero exfiltration and credential exposure, it does not meet yellow or orange trigger thresholds.

5Second-Pass Verification

[ok] 『无网络外发/无凭证读取』结论 — analyze.py imports 仅 pandas/matplotlib.pyplot/seaborn/pathlib；全仓（排除 .git）关键词扫描 requests|urllib|socket|http|subprocess|os.system|getenv|os.environ 零命中；URL 仅出现在 README/resources/README 的徽章与社区链接。
[discrepancy] 日期列判据与文档不符（实测） — SKILL.md Notes 写 'Automatically detects date columns (columns containing \'date\' in name)'，实际代码是 "'date' in c.lower() or 'time' in c.lower()"。实测在仓库自带演示集上命中 **customer_lifetime_value**（'life**time**' 含 'time'），随后 pd.to_datetime 把 2400 一类整数按纳秒解释成 1970 年，输出 'Date range: 1970-01-01 … to 1970-01-01 …'、'Span: 0 days'，并据此画出无意义的时间序列图（实测 pandas 3.0.5；该行为与 pandas 2.x 一致：to_datetime(int) 亦按纳秒）。即 README 宣称该演示集『Perfect for showcasing time-series analysis』，实际触发的是误判列。
[discrepancy] 『自适应/智能识别行业数据类型』的能力声明 — SKILL.md 列出 Sales/E-commerce、Customer、Financial、Operational、Survey 五类数据的『相应分析』，README 写 'Automatically detects data type (sales, customer, financial, survey, etc.) and applies relevant analysis'；但 analyze.py 没有任何行业或语义分类代码，全部行为由 dtype 与列名子串决定（number→describe/corr/hist，object→value_counts/barh，含 date|time→区间）。属**能力夸大**：它确实会因列构成不同而产出不同报告，但没有『识别数据类型』这一步。
[discrepancy] Notes『Generates visualizations only when date columns are present』 — 实际仅 time_series_analysis.png 需要日期列；distributions.png 只需数值列、categorical_distributions.png 只需分类列、correlation_heatmap.png 只需 ≥2 数值列。该句把条件说得比实现更严，属文档描述偏差（本身不影响使用）。
[discrepancy] SKILL.md『Files』清单与仓库实际文件 — SKILL.md 只列 analyze.py、requirements.txt、resources/sample.csv、resources/README.md；仓库另有 examples/showcase_financial_pl_data.csv（README 重点宣传的演示集）与 csv-data-summarizer.zip（分发包）。清单不完整。
[ok] 分发包与源码一致性 — zip 内 analyze.py（sha256 fa217db952541d59…）、SKILL.md（1599338ac2db52ba…）、requirements.txt（3787888dbf39eb94…）与工作区同名文件哈希完全相同；zip 结构为 SKILL.md/analyze.py/requirements.txt/resources/{sample.csv,README.md}。
[ok] 依赖声明三重一致 — SKILL.md frontmatter metadata（python>=3.8, pandas>=2.0.0, matplotlib>=3.7.0, seaborn>=0.12.0）与 requirements.txt 三行、README Technical Details 段逐项吻合；实际运行时三项库均被 import（seaborn 用于 heatmap，matplotlib 用于绘图）。
[ok] 写盘行为与覆盖风险 — 四处 plt.savefig 全部使用相对文件名（correlation_heatmap.png / time_series_analysis.png / distributions.png / categorical_distributions.png）+ dpi=150 + plt.close()；不创建目录、不检查同名文件、不接受输出路径参数——已在 file_writes 与 risks 记入。

6Verdict & Scenarios

Zero friction: drop a CSV into the conversation to receive a complete report and multiple charts without answering any questions—eliminating the primary friction point of analytical assistants.

Minimalist, readable, and hackable implementation: a single 182-line file with no abstraction layers; anyone familiar with pandas can add domain-specific calculations or modify charts within minutes (modification steps are clearly outlined in resources/README).

Stable and complete report structure: fixed sequence across eight sections (dimensions, types, missing values, descriptive statistics, correlations, categorical frequencies, time intervals, chart inventory), making it easy to skim or parse programmatically.

Disciplined conditional output: plots are generated only when supported by the underlying data (no time series without dates, no heatmap without multiple numeric columns), preventing junk artifacts.

Entirely local execution with zero outbound traffic or credentials: well-suited for sensitive spreadsheets containing customer or financial fields (data never leaves the machine).

Best for: Ideal for: quickly surveying an unfamiliar CSV (columns, types, missing values, scale, preliminary correlations, and distributions); non-technical users seeking 'instant reports upon upload'; users processing sensitive tables (customer/financial data) locally without exfiltrating data; and developers wanting a minimal, easily modifiable single-file template for custom domain analysis scripts.

Not for: Not suitable for: domain-specific business analytics (no industry logic included; requires manual extension); scenarios requiring rigorous statistical/causal inferences or anomaly testing (provides purely descriptive statistics without testing or disclaimers); massive or streaming datasets (loads entire file into memory without chunking); automated pipelines requiring deterministic output directories or file overwrite protection (filenames are hardcoded); and environments requiring multi-encoding/cross-locale character set robustness (relies on default pd.read_csv encoding assumptions).

Install Agent-Ready

① Verified Mirror 2026-09-15

Option A · Direct DownloadDownload csv-data-summarizer.tar.gz

sha256: 812ec659a4b533a8…

Option B · Copy Guide for Agent

install.json

Agents automatically download and verify mirrors from this JSON guide.

② Upstream GitHub

GitHub Repo ↗

Pinned to commit 9b3affd270.