11 Skills + 9 Agents: How AskTable Packages Data Analyst Capabilities into AI

What's the gap between an excellent data analyst and a general AI?

Not knowledge volume, not computing power, certainly not who "understands data more".

The gap is methodology.

When a senior data analyst sees sales dropped 20%, they don't vaguely answer "there could be many reasons". They本能地:

1. •First look at baseline - how much deviation from normal range?
2. •Then decompose - which region, category, time period contributed most to the difference?
3. •Then attribute - which factor is driving: price, traffic, conversion rate?
4. •Finally give suggestions - what should be done next.

This process is his years of practice accumulated "tacit knowledge" - he knows what to look at first, what next, which metrics are key, which noise can be ignored.

General AI doesn't lack data analysis knowledge, but executable methodology frameworks. It knows many analysis techniques but doesn't know "in this business scenario, what sequence, standards, granularity should be used for analysis".

What AskTable does is extract this tacit knowledge, package it into capabilities AI can understand and execute.

This article doesn't talk about "what's included" (if you want to know AskTable's built-in skills and agents, see this article), but "how it's done" - from expert experience to systematized capabilities, from customization to standardization, how AskTable completed this "capability packaging" step by step.

---

I. From Tacit to Explicit: How Skills Are "Packaged"

1.1 What Exactly Is a Data Analyst's "Tacit Knowledge"

Every senior data analyst has their own analysis framework, but they often can't articulate it clearly. If you ask "how do you analyze sales decline", they might answer "just look at the data".

But actually, their analysis process is highly structured:

Discover anomaly → Confirm baseline → Multi-dimensional decomposition → Locate core drivers → Exclude occasional factors → Give suggestions

This process contains three levels of knowledge:

• •Analysis logic: What to look at first, what next, what to skip under which circumstances
• •Judgment standards: What is "anomalous"? 5% or 15% deviation? What statistical methods to use?
• •Expression standards: Conclusions first, data supports, suggestions actionable, reports have fixed structure

These are data analyst's "tacit knowledge" - they exist in experience but never systematically coded.

Three Characteristics of Tacit Knowledge

Why is this knowledge hard to pass on? Because they have three typical characteristics:

Characteristic 1: Intuitiveness

Senior analysts can "see at a glance" when something's wrong. But this isn't magic - it's pattern recognition ability formed from extensive practice. Their brain stores hundreds of "data pattern → business reason" mappings, matching automatically when seeing data.

Characteristic 2: Contextualization

Same analysis action, completely different execution in different industries, companies, periods. For example "sales decline" analysis:

• •Retail industry: Look at stores first, then categories, then time periods
• •SaaS industry: Look at renewal rate first, then new customers, then churn
• •Manufacturing: Look at order volume first, then capacity, then delivery

General AI lacks this contextual judgment ability.

Characteristic 3: Non-standardized

Different analysts' report styles vary enormously: some like data first, some conclusions first; some focus on details, some on big picture. But all excellent analysts share one thing - their reports let non-data people quickly understand and take action.

1.2 AskTable's Packaging Approach

AskTable's approach: transform tacit knowledge into explicit instructions, transform explicit instructions into reusable skills.

Specifically, AskTable deconstructs each skill into four elements:

Each skill is essentially a structured system instruction telling AI Agent:

"When users raise such questions, answer following this flow, with these standards, in this format."

Why Instructions Not Code?

In traditional software systems, analysis capabilities are implemented through code. For anomaly detection, you might need hundreds of lines of Python code handling data cleaning, statistical calculation, visualization.

But in AI Agent era, capability packaging fundamentally changed:

Code implementation: Data cleaning → Statistical calculation → Visualization → Result output (engineer maintained)
Instruction implementation: Tell AI "how to analyze" → AI executes autonomously (industry expert maintained)

Instruction packaging's core advantages:

• •Understandable: Industry experts can directly read and modify instructions, no programming needed
• •Iterable: Adjusting analysis methods only needs changing instructions, not code, testing, release
• •Reusable: Same instructions can be shared across different projects, different agents
• •Composable: Multiple skill instructions can stack, forming more complex analysis capabilities

1.3 The Cycle from Instructions to Code to Instructions

AskTable's skill development process follows a unique cycle:

Industry expert experience → Structured analysis framework → AI-executable instructions → Actual effect validation → Instruction iteration optimization → Standardized skill release

This process's key: skills aren't code but AI-understandable instructions. This means:

• •Non-technical personnel can participate in skill design and optimization
• •Skills can be adjusted anytime without recompiling and deploying
• •Same skills can be reused across different projects, different agents

This is also AskTable Skill system's core design philosophy - defining capabilities through instructions not code. For Skill system's technical architecture details, refer to this deep dive.

---

II. 11 Skills Overview: A Data Analyst's Toolbox

AskTable has 11 built-in skills covering a professional data analyst's core capabilities in daily work.

From "capability packaging" perspective, interpret how each skill transforms analysis methodology into AI-executable instructions.

1. Anomaly Detection - From "Seeing Something's Wrong at a Glance" to Algorithms

How analysts do it: Old analysts看一眼趋势图就知道 "this point is off". Their judgment comes from intuitive understanding of historical fluctuations.

How AskTable packages it:

• •Auto-calculate historical baseline (moving average, percentiles, etc.)
• •Set dynamic thresholds (not fixed values but dynamic ranges based on volatility)
• •Mark anomaly time points, give deviation percentages
• •Analyze possible business reasons, recommend drill-down dimensions

Core methodology: Anomaly isn't about absolute values but "deviating from normal patterns".

2. Prediction Trend - From "Gut Feeling Estimation" to Quantitative Prediction

How analysts do it: Combine historical trends, seasonality, known business events to give a "大概" prediction.

How AskTable packages it:

• •Identify trends, seasonality, cycle patterns
• •Give point predictions and confidence intervals (not just one number but a range)
• •Warn about interference from anomalous values on predictions

Core methodology: Prediction isn't giving one number but giving confidence level and risk boundaries.

3. Drill-Down Metrics - From "Peeling the Onion Layer by Layer" to Automated Process

How analysts do it: From total sales → by region → by category → by time period, layer by layer finding problem root causes.

How AskTable packages it:

• •Auto-decompose by time/region/category/person dimensions
• •Find dimensions with largest contribution differences (not all dimensions, focus on key ones)
• •Progress layer by layer until finding specific actionable problem points

Core methodology: Drill-down isn't aimless exploration but "finding largest contribution difference source".

4. Comparative Analysis - From "Compared with Whom, Comparing What" to Standard Action

How analysts do it: Naturally think of QoQ, YoY, comparing with competitors, with targets.

How AskTable packages it:

• •Time comparison (QoQ/YoY/fixed base)
• •Spatial comparison (region/store/department ranking)
• •Plan comparison (actual vs target/budget)
• •Identify key drivers of differences

Core methodology: Comparison isn't listing data but "finding meaningful reference systems".

5. Attribution Analysis - From "Whose Credit Is It Anyway" to Quantitative Decomposition

How analysts do it: Revenue grew, was it price increase or volume increase? Profit improved, was it cost reduction or efficiency improvement?

How AskTable packages it:

• •Decompose each factor's incremental contribution
• •Calculate each factor's elasticity coefficient
• •Identify internal/external/occasional factor impacts

Core methodology: Attribution isn't listing reasons but "quantifying each factor's contribution degree".

6. Stress Testing - From "What Happens in Worst Case" to Scenario Simulation

How analysts do it: When boss asks "what if revenue drops 30%", analyst quickly deduces in head.

How AskTable packages it:

• •Design optimistic/baseline/pessimistic three scenarios
• •Find most critical impact factors
• •Evaluate risk exposure, give response plans

Core methodology: Stress testing isn't scaring people but "proactively identifying risk boundaries".

7. Cycle Analysis - From "There's Always a Pattern Feeling" to Data Verification

How analysts do it: "Sales always drop on weekends", "our peak is Q3" - these are experiential talks.

How AskTable packages it:

• •Identify short/medium/long-term cycles
• •Calculate seasonality indices
• •Trend analysis after removing seasonality

Core methodology: Cycle analysis's value is "distinguishing real trends from false fluctuations".

8. Report Orchestration - From "Writing Reports" to Auto Assembly

How analysts do it: First put core conclusions, then key data, then detailed analysis and suggestions - each report has fixed routine.

How AskTable packages it:

• •Auto-organize content structure by analysis purpose
• •Fill analysis results and visualization charts
• •Generate complete hierarchy of summary, body, appendix

Core methodology: Reports aren't data piling but "structured narrative with logic".

9. Metric Interpretation - From "What Does This Data Mean" to Business Language

How analysts do it: Translate "conversion rate dropped 3.2 percentage points" to "out of every 100 visitors, 3 fewer transactions".

How AskTable packages it:

• •Conclusions first, data follows
• •Explain complex metrics with analogies
• •Enhance data impact with comparisons

Core methodology: Good metric interpretation lets "non-data people understand what data is saying".

10. Data Quality Detection - From "Is This Data Right" to Auto Review

How analysts do it: When getting data, first check for missing, duplicates, obviously illogical values.

How AskTable packages it:

• •Auto-detect missing values, anomalous values, duplicates
• •Validate data consistency and completeness
• •Generate data quality reports

Core methodology: Data quality check is prerequisite for analysis, not optional.

11. Business Language Generation - From "Statistical Results" to "Human Language"

How analysts do it: Translate "p-value < 0.05" to "this difference is statistically significant, not random fluctuation".

How AskTable packages it:

• •Translate technical descriptions into business language
• •Maintain conclusion accuracy while enhancing readability
• •Adjust expression depth by audience

Core methodology: Data analysis value isn't in analysis itself but "making decision-makers understand and take action".

---

III. Agents: Giving Skills Roles and Business Context

Skills are toolboxes, but tools don't work on their own. Agent's essence is equipping these tools with roles, business knowledge, and work habits.

3.1 Skills vs Agents: From Capabilities to Roles

One sentence to distinguish:

• •Skills are "what it can do" - anomaly detection, attribution analysis, report orchestration
• •Agents are "who you are" - retail operations analyst, e-commerce data monitor, financial data analyst

Agent = Skill combination + Role setting + Business knowledge + Work habits

3.2 9 Industry Agents' "Role Formulas"

AskTable's built-in 9 agents, each a validated "role formula":

3.3 Agents Are Not Templates but "Virtual Colleagues"

AskTable defines agents not as "preset configurations" but "virtual colleagues you can work with".

This means:

• •Clear role boundaries (won't cross boundaries to do work not belonging to them)
• •Work habits (proactively push daily reports, anomaly alerts, regular summaries)
• •Communication style (executive assistant concise and efficient, financial analyst rigorous)
• •Learning ability (accumulate preferences and context through memory system)

---

IV. Technical Implementation: How AskTable Packages These Capabilities

4.1 Core Architecture: Skill System

AskTable's capability packaging is built on Skill system. This system's key design decision: skills are data, not code.

class SkillModel(Base):
    id: UUID                          # Skill unique identifier
    project_id: str                   # Project affiliation
    name: str                          # Skill name
    description: str                  # Skill description (for Agent understanding)
    content: str                       # Skill instruction content
    created_at: datetime
    modified_at: datetime

This simple data model behind is a complete capability packaging system:

• •content field stores plain text instructions, not code logic
• •project_id implements project-level isolation, different projects can have different skill libraries
• •Supports complete CRUD operations, skills can be added/deleted/modified anytime

4.2 Three-Layer Loading Mechanism

Skills aren't all stuffed into Agent. AskTable uses three-layer loading mechanism ensuring only necessary skills loaded per conversation:

This design's core value:

• •Context精简: Don't stuff all skills into every request, reduce reasoning cost
• •On-demand activation: Agent judges what capabilities needed during conversation, dynamically call
• •Flexible switching: Same Agent can load different skills based on different conversations

4.3 activate_skill: Skill's "Switch"

async def activate_skill(skill_name: str) -> str:
    """Activate specified skill, return complete skill instructions."""
    if skill_name not in skill_map:
        return f"Skill '{skill_name}' does not exist"
    return skill_map[skill_name]

This seemingly simple tool is the core hub of capability packaging:

1. •Agent reasoning: Understand user intent, judge what professional capabilities needed
2. •Call activate_skill: Load corresponding skill instructions on-demand
3. •Get skill instructions: Obtain structured analysis methodology
4. •Execute analysis: Complete analysis following skill instruction flow and standards

---

V. Custom Skills: Inject Your Industry Knowledge into AI

AskTable's built-in 11 skills and 9 agents cover common scenarios, but every industry and enterprise has its own unique methodology.

AskTable's real value isn't what it provides, but what you can add to it.

5.1 When to Create Custom Skills

Consider custom Skills in these scenarios:

• •Industry-specific metrics: Your industry has unique analysis metrics or calculation methods
• •Enterprise-exclusive processes: Your company has its own analysis processes, report templates, approval chains
• •Business experience accumulation: Senior employees' analysis methods, judgment standards, decision logic
• •Compliance requirements: Industry regulatory required specific check items and report formats

5.2 Custom Skill Best Practices

Suggestion 1: Use "you" to define role

✅ Good: "You are a retail industry analyst with 10 years of experience..."
❌ Bad: "This skill is used to analyze sales data..."

Suggestion 2: Use numbers to define flow

✅ Good:
1. First step
2. Second step
3. Third step

❌ Bad:
"Do this first, then that, then look at..."

Suggestion 3: Use specific numbers to define standards

✅ Good: "Deviation exceeding 15% from baseline viewed as significant anomaly"
❌ Bad: "Relatively large deviation needs special attention"

---

Summary: AskTable Did a "Translation" from Expert Experience to AI Capabilities

AskTable's essence isn't creating a "smarter AI" but building a bridge - translating data analyst's tacit experience into AI-executable structured instructions.

This "translation" key steps:

1. •Extract: Refine analysis frameworks, judgment standards, output specifications from industry expert experience
2. •Package: Transform methodology into Skill instructions, store as reusable data objects
3. •Compose: Assemble skills + roles + business knowledge into industry agents
4. •Extend: Through custom Skills, let each enterprise inject its own exclusive knowledge
5. •Evolve: Agents continuously learn during use, capabilities continuously improve

Three Layers of Capability Packaging

AskTable's capability packaging can be understood as three progressive layers:

The ultimate result:

• •Enterprises no longer rely on "few people" data analysis capabilities
• •Every team can have "virtual colleague"-level professional support
• •Expert experience no longer lost with personnel turnover but deposited as organizational assets
• •Newcomers can stand on shoulders of "virtual seniors" to get up to speed quickly

Data analysis's future isn't training more analysts, but making analyst capabilities触手可及 tools for everyone.

What AskTable is doing is making this process from "custom development" to "standard configuration".