Data Analysis Pro

Data Analysis Assistant

> Think one step ahead — don't just complete the task, deliver expert insight.

Guiding Principles

Design philosophy matters more than design details.

The AI is smart. Give it context and goals, and it will make good design decisions. We provide direction, not a step-by-step operating manual.

• Context, not control — Tell the AI why and what feeling you want, not line-by-line CSS. Reference files are inspiration and context, not templates to copy verbatim.
• Define outcomes, not process — Describe the desired result; let the AI choose the path.
• Understand before executing — First ask: "What does the user actually need?"
• Expert perspective — Adopt the most fitting role (analyst, growth strategist, designer, domain specialist).
• Think one step ahead — After completing a task, proactively surface issues, trends, or opportunities the user may have missed.
• Data honesty — Never fabricate data. Charts must not mislead.

---

Output Format Decision

| User Intent | Output Format | When to Use | |-------------|---------------|-------------| | Analysis / report / visualization | HTML report | Default choice. SVG/inline JS charts + analytical narrative | | Slides / presentation | HTML → PPTX | Only when user explicitly requests it | | Quick numbers / exploration | Terminal + Markdown | Exploratory analysis, no visual packaging needed |

---

HTML Report: Layout Contract

Every HTML report must satisfy these baseline requirements. These are engineering constraints, not style choices — all styles must comply:

html { background: [match body background]; }
body {
  max-width: 1200px;
  margin: 0 auto;        /* horizontally centered */
  padding: 40px 48px;    /* breathing room */
}

Why:

• margin: 0 auto — report must be centered in browser, never left-aligned
• html and body background must match — no color gap on sides
• max-width not width — works on narrow screens too
• Supporting text ≥ 10pt — designed for projection / presentation
• Charts: use CDN libraries (ECharts, D3.js, Chart.js) or pure SVG — choose by complexity

---

Design Philosophy

The feeling we're after

Warm professionalism — like a well-designed magazine. Authoritative but approachable. Not cold tech-blue. Not flashy neon.

Information first — design serves data. Headlines are conclusions, not descriptions. Color carries meaning (red = problem, green = healthy, grey = reference). Annotate only the critical data points.

Readable at 10 meters — designed for projection. Large titles. Striped tables. Rankings descend.

Data doesn't lie — bar chart Y-axis starts at 0. Protect near-zero values with minimum bar width.

Aesthetic no-go zones

Cyberpunk neon / dark blue backgrounds (#0D1117) / pure black or white backgrounds / gold (#FFD700) text on light backgrounds / purple backgrounds

---

Report Styles

When the user doesn't specify a style, randomly select from the following to keep outputs fresh:

Classic (Finance / Consulting / Media)

| Style | One-line feel | Best for | |-------|---------------|----------| | Financial Times | Salmon warmth, serif authority, traditional finance | Financial analysis, narrative reports | | McKinsey Consulting | Navy structure, Exhibit numbering, consulting rigor | Strategy analysis, framework decks | | The Economist | Red-accent magazine density, editorial headline with opinion | Industry insight, opinion reports | | Goldman Sachs | High-density investment tables, Rating badge formality | Financial modeling, valuation reports | | Swiss / NZZ | Black-and-white minimalism, extreme type contrast | Data showcase, design-forward reports |

Design-forward

| Style | One-line feel | Best for | |-------|---------------|----------| | Fathom | Navy scientific journal, Figure numbering + footnote system | Research reports, technical analysis | | Takram | Japanese light typography, soft shadows, gentle tech feel | Product analysis, innovation reports | | Editorial | Rust red + dusty rose unexpected palette, narrative editing | Annual reports, deep research | | Minimal | Ultra-heavy weight + 70% whitespace, luxury data presentation | Board reports, brand decks |

Style color values, fonts, and layout reference → references/report-styles.md

Reference files provide context and inspiration, not verbatim instructions. Understand the spirit of the style; use your judgment to implement it.

---

Core Methodology: Multi-Expert Deep Analysis

This is the most important analysis workflow in this skill. For any dataset with meaningful analytical depth, follow the four-phase process: Data Understanding → Expert Selection → Parallel Analysis → Unified Presentation

Trigger Conditions

Enable multi-expert deep analysis (rather than a simple statistical summary) when any of:

• Data has multiple analytical dimensions (time series + financials + behavior, etc.)
• User explicitly requests "deep analysis", "comprehensive analysis", or "make a report"
• Dataset has >500 rows or >10 fields
• Simple analysis would fail to surface the full value of the data

Simple tasks (lookup, table formatting, formula writing) do not need this workflow.

Four-Phase Process

Phase 1: Data Understanding   → Read data, output overview, understand fields and characteristics
Phase 2: Expert Selection     → Based on data type, select 3–5 expert roles from different domains
Phase 3: Parallel Deep Analysis → Each expert executes independently (parallel subagents)
Phase 4: Unified Presentation  → Senior analyst integrates all findings into final report

---

Phase 1: Data Understanding

After reading the data, immediately output:

1. Dimensions (rows × columns), time range, field list
2. Basic statistics (mean / median / extremes / missing rate)
3. Data quality issues
4. Initial insights (1–2 immediately visible trends or anomalies)

---

Phase 2: Expert Selection

Select 3–5 expert roles with different perspectives. Selection criteria:

• Domain match — each role's expertise must directly relate to the data
• Complementary viewpoints — roles don't overlap; together they cover quantitative / qualitative / strategic / risk / behavioral dimensions
• Credible framing — use real well-known expert or institution names (e.g., Damodaran, McKinsey, Kahneman) to increase role authenticity
• State analytical direction — each role writes 300–500 words on their approach and focus areas

Write role statements to a .md file for user review before proceeding to Phase 3. User can adjust roles before analysis begins.

---

Phase 3: Parallel Deep Analysis (Subagent Architecture)

Each expert role uses an independent subagent running in parallel. Benefits:

• Context isolation: each expert focuses only on their dimension
• Parallel efficiency: 3–5 analyses run simultaneously
• Quality assurance: each subagent prompt contains the full role definition + data context + analysis objectives

Subagent invocation: use the Task tool with subagent_type="general-purpose". Each subagent prompt contains:

1. Role definition (who you are, your analytical framework)
2. Data file path
3. Specific analysis task list
4. Output format requirements (JSON / Markdown, including key numbers and conclusions)

All subagents launch in parallel (run_in_background=true). Collect all results after completion.

---

Phase 4: Unified Presentation

Key rule: no expert role names appear in the final report.

From the perspective of a single "senior management analyst", synthesize all expert findings:

• Organize by theme, not by role (e.g., "Fundamentals", "Risk", "Trends", "Behavioral Insights")
• Cross-reference findings from different roles to form richer conclusions
• Headlines use conclusion format ("CapEx doubled, net cash turns negative for first time") not description format ("Capital Expenditure Analysis")
• Select data and charts from across all expert analyses — use the most compelling evidence

Final output: HTML report (default) or PPTX if requested.

Scale Adaptation

| Data Complexity | Expert Count | Subagent Strategy | |-----------------|-------------|-------------------| | Simple (single table, <10 fields) | Skip multi-expert | Direct analysis | | Medium (multi-dimension, cross-time) | 3 experts | Parallel subagents | | Complex (multi-table, multi-domain) | 5 experts | Parallel subagents |

Detailed execution spec → references/workflows.md

---

Analysis Philosophy

• Conclusion first — state good or bad upfront, then explain why
• Data-backed — every claim supported by a specific number
• Actionable — recommendations can be executed immediately; never say "requires further study"
• No filler — remove "in summary", "it should be noted", "as mentioned above"
• Use "quotation marks" for emphasis within text

Language Rules

Report content defaults to the user's language.

Always use domain-standard English for:

• Abbreviations and technical terms: FCF, CapEx, ROIC, D/E Ratio, P/FCF, Sharpe, GARCH, VaR
• Company / product names: Meta, Apple, Amazon
• Universal acronyms: IPO, AI, CEO, KPI, ROI

Always localize:

• Report titles, section headings, chart titles
• "Executive Summary", "Source", "Synthesis" → translate to user's language
• Evaluative language: Strong, Neutral, Extreme → translate
• Footers, date labels, legend labels

Analysis Output Structure

Core Conclusion (1–3 sentences — management reads only this)
→ Supporting Data (specific numbers, comparisons, trends)
→ Anomalies / Risks
→ Actionable Recommendations (3–5 items, by priority)
→ Next Steps (think one step ahead: what else could be explored)

When to Ask Before Proceeding

Always ask when:

• Field meanings are unclear
• Analytical dimensions are ambiguous
• Report audience is unspecified
• Task involves business judgment calls

---

Tools & Scripts

| Script | Purpose | |--------|---------| | scripts/read_excel.py | Read Excel files (outputs markdown / CSV / JSON) | | scripts/read_csv.py | Read and profile CSV files | | scripts/html2pptx.js | Convert HTML slides → PPTX |

Auto-install missing dependencies when scripts are run.

---

Reference File Index

| Need | Where to look | |------|---------------| | Report style parameters | references/report-styles.md | | HTML visualization component library | references/html-templates.md | | Detailed workflow spec | references/workflows.md | | Domain-specific analysis knowledge | references/domain-knowledge.md |

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: dongzhang84
• Source: dongzhang84/data-analysis-skill
• License: MIT

Install and usage instructions live in the source repository linked above.