CRITICAL - This skill orchestrates a multi-step analysis flow. Execute in the order of
references/steps/; do not skip steps or fabricate data. CRITICAL - Step 01 is a mandatory gate. The analysis target (short drama name) and analysis time range must be explicitly confirmed by the user before data collection begins. Never infer the drama name automatically when it is missing. CRITICAL - Use onlyae-clifor all data operations. Command patterns are inreferences/adapters/ae-cli.md. Do not guess command names, flags, JSON payloads, project_id, event names, or property names. CRITICAL -references/config/event_mapping.yamlis the bridge between business concepts and fields. The standard event/property candidate-name lists live in this file; reuse it to complete the mapping when switching projects (mapping results go to user memory, never written back to this file).
Applicable Scenarios
- Post-launch performance evaluation of a drama
- Locating problematic episodes and their root causes
- Providing data support for content optimization and promotion decisions
- Supports three formats: vertical mini-drama (Douyin/Kuaishou), horizontal short drama (iQiyi/Youku/Tencent), and private-platform drama
Global Conventions
Interaction Checkpoints
The whole flow has two mandatory interaction checkpoints that let the user participate in decisions instead of running through to completion in one go:
| Checkpoint | Location | User Decision |
|---|---|---|
| Checkpoint 1 | After Step 02 data collection | Choose which dimensions participate in scoring |
| Checkpoint 2 | After Step 03 score calculation | Whether to run deep diagnosis, output the report directly, or inspect a specific episode first |
Parallelism Convention
- Step 02 runs queries in three dependency-aware phases (see steps/02_query_data.md and query_patterns.md): Phase A issues all independent synchronous queries in parallel; Phase B runs the paywall-dependent payment funnel and submits the async detail exports; Phase C polls the async artifacts. Queries within a phase run in parallel, across phases in order — never guess a dependency away or serialize independent calls.
- A single dimension query failure does not block the overall flow; the final report marks "this dimension was not collected due to XX and has been skipped."
Language Convention
- Skill instructions, config files, templates, and reference docs are written in English (language-compliance requirement).
- User-visible output MUST be translated to Chinese before being presented: report content, checkpoint summaries, interaction prompts, diagnostic wording, and confirmation dialogs.
- Translate dimension names, grade labels, and suggested actions using the authoritative terms in Chinese Output Glossary so terminology stays consistent across reports.
- Event/property names in the report keep the mapped project-specific actual names (never translate event/property identifiers).
Chinese Output Glossary
Standard Chinese terms for user-facing output. When translating English terms from config files or templates into the report, use these exact terms:
| English term | Chinese title |
|---|---|
| First-Episode Appeal | First-episode appeal |
| Completion Quality | Completion quality |
| In-Episode Pacing | Intra-episode pacing |
| Cliffhanger Effect | Hook effectiveness |
| Inter-Episode Retention | Inter-episode retention |
| Binge Depth | Binge-watching depth |
| Payment Conversion | Paid conversion |
| Heat Trend | Popularity trend |
| User Engagement | User engagement |
| Potential Hit (grade S) | Hit potential |
| Quality Content (grade A) | High-quality content |
| Average (grade B) | Mediocre |
| Multiple Issues (grade C) | Many issues |
| Poor Quality (grade D) | Poor quality |
| Full Promotion | Full-scale promotion |
| Continued Investment | Sustained investment |
| Targeted Optimization | Targeted optimization |
| Key Rectification | Priority rectification |
| Stop Loss | Recommend cutting losses |
| Vertical mini-drama | Vertical micro short drama |
| Horizontal short drama | Horizontal short drama |
| Private-platform drama | Own-platform short drama |
Key Principles
Before analysis, the short drama name and analysis time range must be confirmed first; only after confirmation load config files and call ae-cli.
- In-Episode Pacing degrades gracefully: it needs either a play-progress event with a progress property, or a play-duration property (detail-export fallback). When neither exists, the whole dimension is marked N/A (its weight is reallocated) rather than guessed.
Architecture
drama-quality-assessment/
├── SKILL.md ← Orchestrator (this file)
├── references/
│ ├── adapters/ae-cli.md ← ae-cli command patterns (tool adapter layer)
│ ├── config/
│ │ ├── dimensions.yaml ← 9 dimension definitions, weights
│ │ ├── scoring_rules.yaml ← scoring thresholds, tiered mapping rules
│ │ └── event_mapping.yaml ← standard event/property candidate-name list
│ ├── steps/
│ │ ├── 01_parse_config.md ← business confirmation + technical config
│ │ ├── 02_query_data.md ← data collection (9 dimensions in parallel)
│ │ ├── 03_calculate_score.md ← score calculation + heatmap
│ │ ├── 04_diagnose.md ← AI diagnosis
│ │ └── 05_generate_report.md ← report generation
│ ├── templates/
│ │ └── report_template.md ← report template
│ ├── dimension_details.md ← dimension calculation formulas
│ ├── query_patterns.md ← per-dimension query templates (definition JSON)
│ └── anomaly_patterns.md ← library of 12 anomaly/opportunity patterns
Step Overview
Step 1 Parse Config → Step 2 Data Collection → 🛑Checkpoint 1 → Step 3 Score Calculation → 🛑Checkpoint 2 → Step 4 Diagnosis → Step 5 Report Generation
| Step | File | Depends on | Output |
|---|---|---|---|
| 01 | steps/01_parse_config.md | — | Business/technical config, event mapping, drama filter |
| 02 | steps/02_query_data.md | Step 01 | Raw data for 9 dimensions |
| 🛑 Checkpoint 1 | — | Step 02 | User chooses scoring scope |
| 03 | steps/03_calculate_score.md | Step 02 | Scorecard + heatmap |
| 🛑 Checkpoint 2 | — | Step 03 | User decides whether to diagnose |
| 04 | steps/04_diagnose.md | Step 03 | Diagnostic insights + optimization suggestions |
| 05 | steps/05_generate_report.md | Step 03/04 | Full report |
Core Index
Config
| File | Purpose |
|---|---|
| config/dimensions.yaml | 9 dimension definitions, weights, metric list |
| config/scoring_rules.yaml | Scoring thresholds, tiered linear mapping rules, grade definitions |
| config/event_mapping.yaml | Standard event/property candidate-name list (business-concept ↔ field bridge) |
Reference Docs
| File | Purpose |
|---|---|
| dimension_details.md | Calculation formulas and special handling rules per dimension |
| query_patterns.md | Per-dimension query templates (definition JSON) |
| anomaly_patterns.md | Library of 12 anomaly/opportunity patterns (used by Step 04) |
Output Template
| File | Purpose |
|---|---|
| templates/report_template.md | Full report structure (five sections) |
Tool Adapter
| File | Purpose |
|---|---|
| adapters/ae-cli.md | ae-cli command patterns — load only when ae-cli is available |
9 Evaluation Dimensions
| Dimension | Weight | Key Metrics |
|---|---|---|
| First-Episode Appeal | 15% | First-episode click→play conversion, first-episode completion rate, 1→2 episode jump rate |
| Completion Quality | 10% | Per-episode completion rate, time-segmented completion trend |
| In-Episode Pacing | 10% | Progress retention curve (25%/50%/75%/100%), high-churn positions |
| Cliffhanger Effect | 15% | Episode-end jump rate, 5s cliffhanger jump rate, cliffhanger failure ratio |
| Inter-Episode Retention | 15% | Next-day retention, inter-episode jump rate, churn point distribution |
| Binge Depth | 10% | Binge ≥3 episodes ratio, median binge episodes |
| Payment Conversion | 15% | Payment trigger rate, post-payment watch rate |
| Heat Trend | 5% | Daily average plays, new-user ratio, play peak |
| User Engagement | 5% | Like rate, comment rate, share rate |
Weights can be overridden via conversation, or persisted to references/config/dimensions.yaml.
User Input Requirements
Required: short drama name, TE project ID
Optional (with defaults): analysis time range (default: last 30 days), analysis episode range (default: all), drama type (default: vertical mini-drama), total episode count (auto-inferred), paywall position, custom weights
Weight Customization
- Temporary override: "Set payment conversion weight to 20% and lower heat trend to 3%"
- Persist: "Remember this weight config"
- Restore defaults: "Restore default weights"
Report Output
The report always contains five sections. If the user skips deep diagnosis, sections 3 and 4 remain present with explicit skip notes:
- 📊 Comprehensive Scorecard — total score, grade, per-dimension scores
- 🔥 Per-Episode Heatmap — episode × dimension score matrix, anomalies highlighted
- 🔍 AI Diagnosis — AI-generated data insights and root-cause analysis
- 💡 Optimization Suggestions — concrete action suggestions sorted by priority
- 📝 Data Notes — data source, coverage, missing dimensions, and weight handling
Start Execution
Read and execute references/steps/01_parse_config.md.