Datastory

DataStory — Auto Data Profiler + Narrative Generator

Given any CSV or Excel file, this skill profiles the data, generates an AI narrative via the Anthropic API, and outputs a polished .docx executive report. No manual analysis required.

Scripts

Quick Run (full pipeline)

export ANTHROPIC_API_KEY=sk-ant-...
bash scripts/run_datastory.sh mydata.csv
# Output: datastory_mydata.docx

Step-by-step workflow

Step 1 — Profile the data

python3 scripts/profile_data.py <filepath> [--max-rows 10000] > profile.json

Outputs JSON with:

• shape — rows x columns
• missing_summary — total missing cells and %
• columns[] — per-column: dtype, null %, unique count, and either numeric stats or top_values or datetime range
• sample_rows — first 5 rows

Edge cases:

• Files >10k rows are sampled automatically
• CSV encoding failures fall back to latin-1

Step 2 — Generate narrative via Anthropic API

Call claude-sonnet-4-20250514 with max_tokens: 4000.

System prompt: instruct the model to return ONLY JSON with keys: executive_summary, dataset_overview, key_findings (array of 5), anomalies (array), column_insights (array of {column, insight}), data_quality_score (int 0-100), data_quality_label (Poor/Fair/Good/Excellent), data_quality_assessment, recommended_next_steps (array of 3).

Slim the profile before sending: keep only 3 sample rows and 3 top_values per column. Parse: strip markdown fences, JSON.parse. On failure, extract first {...} block with regex.

Step 3 — Generate .docx report

node scripts/generate_report.js payload.json [output.docx]

payload.json format: { "profile": {...}, "narrative": {...}, "filename": "myfile.csv" }

Report sections:

1. Cover — title, filename, date
2. Executive Summary — callout box + dataset overview
3. Dataset Overview — stats table
4. Key Findings — numbered list
5. Anomalies & Concerns — amber callout (skipped if empty)
6. Column Profiles — full table with name, type, nulls, stats, outliers
7. Column Insights — bulleted insights
8. Data Quality Assessment — score + paragraph
9. Recommended Next Steps — numbered list
10. Appendix: Sample Data — first 5 rows

docx rules: WidthType.DXA only, dual widths on tables, ShadingType.CLEAR, no unicode bullets, no backslash-n in text, outlineLevel on headings for TOC.

Step 4 — Deliver

Copy to /mnt/user-data/outputs/ and call present_files. Name: datastory_<original_filename_without_extension>.docx

Error handling

React Artifact

../app/DataStory.jsx runs this entire pipeline in-browser via the Anthropic API (Claude-in-Claude pattern). Upload a file, get a full interactive report instantly.