AgentSkillsCN

data-analysis

通用助手行为，包括技能路由与文件/工具的使用。

SKILL.md
--- frontmatter
name: data-analysis
description: End-to-end data analysis with DuckDB queries, optional Plotly charts, and artifact-aware summaries.
tools:
  - data_agent_tools
source_skills:
  - data_agent-core
  - data_agent-schema
  - data_agent-sql
  - data_agent-chart
  - data_agent-file
  - data_agent-summary

data-analysis

Overview

Use this skill to analyze tabular data in the workspace using DuckDB tools, optionally produce a Plotly visualization, and deliver a concise, evidence-based summary that references any generated artifacts.

Standards (required)

  • •Summary and insights must be valid Markdown, not raw paragraphs. Use headings (###) and bullet lists (-) so the report renders cleanly.
  • •Every insight must cite concrete evidence from queries (counts, averages, deltas, correlations, percent shares). No speculation.
  • •Name artifacts and charts clearly using Title_Case_With_Underscores and reference them explicitly in the summary.
  • •State limitations if data is incomplete or if a question cannot be answered from available tables.

Default analysis pack (use unless user asks for a quick answer)

  1. •Dataset snapshot: row count, key columns, and basic distribution for the main outcome.
  2. •Segmentation: group by primary category (or tiers) and compute core metrics.
  3. •Drivers: quantify top positive/negative relationships (correlation or ranked deltas).
  4. •Interactions: at least one two-factor combination or tiered interaction (e.g., high/low bins).
  5. •Outliers/variance: highlight spread changes across segments (box plot or variance stats).

Workflow

  1. •

    Scope & schema

    • •Use get_table_schema to inspect only the tables needed for the question.
    • •Note relevant tables, key columns, joins, date fields, and metrics before querying.
  2. •

    Query

    • •Use run_sql_query to answer the question with focused SQL.
    • •Always specify columns (no SELECT *) and include LIMIT 1000.
    • •You may run up to 5 queries total; refine only when needed.
    • •Prefer a segmentation query and at least one interaction/combination query when deeper analysis is requested.
    • •After each query, briefly interpret results and decide whether another refinement is required.
  3. •

    Chart (optional)

    • •If a chart helps, use generate_chart_config on the latest df.
    • •Plotly only. Assign a serializable chart_config and keep logic deterministic.
    • •Plotly outputs will be saved as .plotly.json and .plotly.html in charts/.
    • •Use descriptive chart_title names (Title_Case_With_Underscores).
    • •Do not generate more than 3 charts.
  4. •

    Artifacts (optional)

    • •If artifacts are useful (CSV extracts, HTML previews, Markdown notes), write them to disk using the provided helpers.
    • •Keep artifacts minimal and clearly named; explain why each artifact exists.
  5. •

    Summary

    • •Call generate_summary once after queries (and any chart).
    • •Include tables used, filters, metrics, and artifacts created.
    • •Provide at least two concrete insights (or explain if none exist).
    • •Structure output like:
      • •### Summary with 2-4 bullets
      • •### Key Insights with 3-6 bullets
    • •Mention any artifacts (charts/files) explicitly so the user knows what to open.
    • •Note: generate_summary will also create a report file in reports/ with the analysis steps and outputs.

Guardrails

  • •Order is mandatory: schema -> SQL -> chart (optional) -> artifacts (optional) -> summary.
  • •No direct file I/O or Python reads of raw data files.
  • •Use only the data tools provided (via data_agent_tools).
  • •Explain intent briefly before each tool call; stop after generate_summary succeeds.