4-Phase Orchestration: 5 Universal Agent Skills with YAML-Driven Rules, Composable Components, and Graceful Degradation

When you're hard-coding your 3rd scoring if-else, maybe it's time to ask: can I move the rules into YAML and let the business change config instead of code?

Every Agent developer faces the same dilemma — every business scenario rewrites a similar pipeline:

The skeleton is identical. What changes is only the "content" at each step. Yet every team builds pipelines from scratch.

teleagent-skills offers an answer:

┌─────────────────────────────────────────────────────────────┐
│                    Upper Business Skill                      │
│  (Scoring Engine / Evidence Chain / Data Aggregator / ...)  │
└──────────┬──────────┬──────────┬──────────┬────────────────┘
           │          │          │          │
           ▼          ▼          ▼          ▼
    ┌──────────┐┌──────────┐┌──────────┐┌──────────┐
    │ Phase 1  ││ Phase 2  ││ Phase 3  ││ Phase 4  │
    │ Extract  ││ Analyze  ││ Generate ││ Archive  │
    │          ││          ││          ││          │
    │Info-     ││Data-     ││Report-   ││Archive-  │
    │Extractor ││Analyst   ││Generator ││Manager   │
    └────┬─────┘└────┬─────┘└────┬─────┘└────┬─────┘
         │           │           │           │
         ▼           ▼           ▼           ▼
    ┌─────────────────────────────────────────────────┐
    │          JSON Contract (Structured Data Contract) │
    │   phase1_output.json → phase2_input.json → ...  │
    └─────────────────────────────────────────────────┘

Core idea: each Phase is an independent component, and Phases pass data only through JSON contracts.

{
  "phase": "extract",
  "skill": "scoring-engine",
  "output": {
    "entities": [
      {
        "name": "客户A",
        "type": "enterprise_customer",
        "attributes": {
          "annual_revenue": 50000000,
          "employee_count": 320,
          "industry": "制造"
        }
      }
    ],
    "metadata": {
      "extraction_time": "2026-07-01T10:30:00Z",
      "source": "CRM_API",
      "confidence": 0.92
    }
  },
  "next_phase": "analyze"
}

The traditional approach hard-codes scoring rules:

if customer.revenue > 10000000:
    score += 30
elif customer.revenue > 5000000:
    score += 20

teleagent-skills does it differently — all rules are externalized to YAML:

scoring_engine:
  name: "政企客户商机评分"
  version: "2.1"
  dimensions:
    - id: revenue
      name: "营收规模"
      weight: 0.30
      rules:
        - condition: "attributes.annual_revenue >= 100000000"
          score: 30
          label: "超大型"
        - condition: "attributes.annual_revenue >= 50000000"
          score: 20
          label: "大型"
        - condition: "attributes.annual_revenue >= 10000000"
          score: 10
          label: "中型"
    - id: industry
      name: "行业属性"
      weight: 0.25
      rules:
        - condition: "attributes.industry in ['金融','医疗']"
          score: 25
          label: "高价值行业"
    - id: growth
      name: "增长潜力"
      weight: 0.20
    - id: connectivity
      name: "接入成熟度"
      weight: 0.15
    - id: decision_chain
      name: "决策链清晰度"
      weight: 0.10
  thresholds:
    high: 70
    medium: 40
    low: 0

Business changed? Edit YAML. Need a new dimension? Add YAML. Zero code changes.

All 5 Skills share the same 4-Phase skeleton, but each Skill's Phase behavior differs:

The power of composability: upper-level Skills can chain lower-level Skills on demand. For example, the data query gateway pipeline:

NL2Query(Phase1-2) → Data Aggregator(Phase1-2) → Visualization Renderer(Phase1-3)

One natural language query automatically flows through "understand → query → aggregate → visualize" end to end.

The 4-Phase architecture has built-in 3-tier degradation:

┌─────────────┐     ┌─────────────┐     ┌─────────────┐     ┌─────────────┐
│   Phase 1   │────▶│   Phase 2   │────▶│   Phase 3   │────▶│   Phase 4   │
│   Extract   │     │   Analyze   │     │   Generate  │     │   Archive   │
└──────┬──────┘     └──────┬──────┘     └──────┬──────┘     └──────┬──────┘
       │                   │                   │                   │
       ▼                   ▼                   ▼                   ▼
  ┌──────────┐       ┌──────────┐       ┌──────────┐       ┌──────────┐
  │ Return   │       │ Skip     │       │ Simplify │       │ Local    │
  │ raw      │       │ analysis │       │ template │       │ cache +  │
  │ input +  │       │ mark low │       │ raw data │       │ deferred │
  │ conf=0   │       │ confidence│      │ passthrough│      │ retry    │
  └──────────┘       └──────────┘       └──────────┘       └──────────┘

The core principle of degradation: I'd rather give the user a low-confidence result than crash with an error.

What it is: A multi-dimensional weighted scoring component driven by YAML rule configs.

Typical scenarios: enterprise opportunity scoring, vendor evaluation, customer churn prediction, partner tiering.

Input → Rule matching → Scoring output flow:

{
  "customer": "客户A",
  "total_score": 78,
  "grade": "A级-重点跟进",
  "dimension_breakdown": {
    "revenue": { "score": 20, "max": 30, "label": "大型" },
    "industry": { "score": 15, "max": 25, "label": "中价值行业" },
    "growth": { "score": 20, "max": 20 },
    "connectivity": { "score": 15, "max": 15 },
    "decision_chain": { "score": 8, "max": 10 }
  },
  "recommendation": "建议安排专属客户经理，优先推荐5G专网+云网融合方案"
}

What it is: A multi-source evidence cross-validation component that detects conflicts, evaluates confidence, and pinpoints root causes.

Data Source 1: Customer complaints    ──┐
                                         │     ┌──────────────────┐
Data Source 2: System alert logs      ──┼────▶│  Evidence Chain  │
                                         │     │  Phase2: Analyze │
Data Source 3: SLA monitoring data    ──┘     └────────┬─────────┘
                                                  │
                          ┌───────────────────────┐│
                          │ Cross-validation:      │
                          │ • Complaint: "2hr outage"
                          │ • Alert: "optical attenuation"│
                          │ • SLA: "99.1% availability" │
                          │ Conflict detected:      │
                          │ Complaint vs SLA surface contradiction│
                          │ Root cause="optical attenuation": 0.87│
                          └───────────────────────┘

What it is: A raw data re-processing component supporting validation/cleaning, aggregation, YoY/MoM, and TOP rankings.

Raw query results           Aggregator output
┌──────────────┐         ┌──────────────────────────────┐
│ 300 rows of  │───────▶ │ Monthly summary + YoY/MoM    │
│ detail data  │         │ TOP10 ranking                │
│ (region x    │         │ Anomaly flags (>2σ)          │
│  month)      │         │ Trend direction              │
└──────────────┘         └──────────────────────────────┘

What it is: An automated rendering component that turns structured data into ECharts charts and Dashboards.

Chart type selection is automatic: time-series → line chart, categorical → bar/pie, multi-dimensional → radar.

What it is: A smart natural-language-to-structured-query conversion component.

User input: "华东区上月5G专网新增客户数"

Phase1 Extract:  intent="query", entities=[region="华东", time="上月", metric="5G专网"]
Phase2 Analyze:  Generate SQL + confidence 0.88
Phase3 Generate: Format results
Phase4 Archive:  Record query log

Confidence scoring: When confidence drops below a threshold, the output gets a "low confidence" warning and shows the SQL for human review.

Key insight: All 5 Skills adapt to different industries purely through YAML rule configs. The scoring engine code logic is identical — only the YAML files differ.

git clone https://github.com/yuzhaopeng-up/teleagent-skills.git
cd teleagent-skills

cp skills/scoring-engine/config/scoring_rules.yaml my_rules.yaml

License: Apache 2.0

Stop building pipelines from scratch. The 4-Phase skeleton is ready — you just need to write YAML.

Star teleagent-skills and let's standardize Agent Skills together.