9-Step Workflow Hands-On
Step by step — from data preparation to report review, run the full data analysis pipeline using a single sales dataset to produce a complete analysis report on the spot.
Main case: a single 2026Q1-Sales.xlsx (202 rows of sales records) → complete analysis report. One dataset throughout, report finished by step 9.
Course Learning Assistant
First check whether SkillHub is installed. If not, follow https://skillhub.cn/install/skillhub.md to install SkillHub (CLI only), then install the 通用数据分析流程 skill. If already installed, directly install the 通用数据分析流程 skill.
9-Step Overview
| Step | What It Does | Who's Responsible |
|---|---|---|
| Step 1 | Prepare data: verify, desensitize | You |
| Step 2 | Set background: 6 background elements | You |
| Step 3 | Data profiling: survey, don't analyze | AI |
| Step 4 | Define dimensions: choose slice angles | You |
| Step 5 | Calculate metrics: definitions must be clear | AI |
| Step 6 | Generate charts: pick the right chart types | AI |
| Step 7 | Interpret: findings must include numbers | AI draft + You |
| Step 8 | Build report: one-page dashboard | AI |
| Step 9 | Review: spot-check by hand | You |
Step 1: Prepare Data
Spend 30 seconds verifying the data: where it came from, what time period it covers, how many rows. If it contains sensitive info (phone numbers, ID numbers), desensitize first.
Never feed classified or sensitive information to AI. Desensitize before input, or don't input at all. 4 desensitization methods: masking, replacement, generalization, deletion. Delete sensitive columns that aren't needed for analysis.
Step 2: Set Background
Use the 6 background elements to describe the data clearly so AI doesn't guess:
- Business: Sales detail
- Time: Q1 2026
- Fields: Date, salesperson, product, quantity, amount, province, customer type
- Definition: Tax-inclusive, unique order count
- Question: Regional and product sales performance
- Target: Compare to monthly targets
Step 3: Data Profiling (Prompt A)
Survey first, analyze later. Report only, no analysis:
Don't rush to analyze. First, scan this table and report:
How many columns, what each column means, date range, total rows,
any null values, duplicate records, or anomalous numbers,
whether date formats are consistent.
Report findings only — don't start analyzing yet.Step 4: Define Dimensions
Dimensions = business sense. AI will list 6 options; you choose 4:
| AI-Suggested Dimension | Your Choice | Reason |
|---|---|---|
| By month | Yes | See trends |
| By product | Yes | See rankings |
| By province | Yes | See regions |
| By customer type | Yes | Enterprise vs. individual |
| By salesperson | No | Not evaluating individuals this time |
| By weekday | No | Not enough data |
Few and precise > many and broad. You order from the menu, not AI.
Step 5: Calculate Metrics (Part of Prompt B)
The definition trap is the easiest to fall into: "order count" — by unique order ID or by row count? Target values cannot be fabricated by AI; they must come from the target table.
Step 6: Generate Charts (Part of Prompt B)
4 charts: monthly trend line, product Top10 bar, province proportion donut, enterprise vs. individual comparison.
Chart selection matrix: trends → line, rankings → bar, proportions → donut, comparisons → grouped bar. Be specific in your requirements: colors, titles, axes, legends, units, data source.
Step 7: Interpret
Findings must include numbers, and distinguish facts from speculation:
- Fact: March sales dropped 15% month-over-month (3.71M → 3.15M)
- Speculation: May be related to post-holiday demand decline (to be verified)
Step 8: Build Report (Prompt C)
Based on the analysis results above:
1. Distill the 3 most important findings, written as "finding → supporting numbers → possible cause."
Each finding must point to specific numbers — no vague statements like "overall positive trend."
2. Consolidate into an A4 one-page management dashboard + 300-word summary for management:
Highlight March decline, best-selling products and regions, 3 improvement recommendations.
3. Propose 3 immediately actionable improvement measures, ranked by priority.
For each recommendation, specify: which data supports it, who is responsible, deadline, and how to measure effectiveness.Step 9: Review
Spot-check by hand: pick 3 orders and manually sum them — does it match what AI reported? The anomalous value 99999 should be flagged "to be verified," not "violation."
AI will fabricate numbers with full confidence. Numbers, dates, and names must be verified manually — when AI is wrong, it sounds just as assertive.
Full Prompt B: Sales Analysis Full Workflow
This is the complete prompt for steps 5-6, including 6 background elements + definitions + output format:
This is the company's Q1 2026 sales detail (the file contains sales
detail and monthly targets — two tables). Fields include date,
salesperson, product, quantity, sales amount, province, customer type.
I want to see Q1 overall performance and multi-dimensional breakdown:
1. Total sales, total order count (by unique order ID), average order value
2. Monthly sales and achievement rate, compared to the monthly target table
3. Enterprise vs. individual customer average order value comparison
Definition: amounts are tax-inclusive, order count by unique order ID,
not by row count.
Output: HTML report with monthly trend line chart, product Top10 bar chart,
province proportion donut chart, enterprise vs. individual comparison chart;
corporate blue primary color, full titles, data source noted under each chart.Human-AI Division of Labor Recap
| Your 4 Steps | AI's 5 Steps |
|---|---|
| Step 1 Prepare data | Step 3 Data profiling |
| Step 2 Set background | Step 5 Calculate metrics |
| Step 4 Define dimensions | Step 6 Generate charts |
| Step 9 Review | Step 7 Interpret |
| — | Step 8 Build report |
One-line recap: prepare data → describe clearly → survey first → pick angles → define metrics → draw right charts → speak plainly → build report → you decide.
Prompt Formula & 6 Background Elements
One universal formula (Role + Task + Background + Format + Constraint) plus 6 data-specific background elements to stop AI from guessing and reduce rework.
Self-Practice Templates & Pitfall Guide
8 timeline-ordered prompt templates (data exploration → sales analysis → expense analysis → funnel → survey → chart styling → action recommendations → QA checklist), plus 3 general red lines + 4 data-analysis-specific red lines.
Tutorials