What does AI Story Vault — Self-Evolving Hit Story Think Tank do?
🏷️ Skill Classification Tags
Category: [General Cross-Type] (Topic selection / Hit research / Script assistance layer, does not itself decide live-action or comic style, serving both realistic and dynamic comic story shorts)
- 🎬 Typical Task Scenarios: Research "what topics are hot right now" and search the internet for the latest hit data; deconstruct why a certain AI story short went viral; analyze a hot creator's style; find benchmark cases; write a tear-jerker / family / suspense short script based on the latest hit patterns; track the latest AI video trends.
- 🚫 Non-Applicable Scenarios: ① Systematic adaptation/upgrade of an existing novel/old script → use
blockbuster-script-director; ② Breaking a script into scenes/characters/props asset images → usescript-asset-master; ③ Directly generating/editing videos → useai-drama-master-suite/jimeng-director/jianying-editor.- 🔗 Upstream/Downstream / Recommended Pairing: This skill (online topic research / hit analysis / hook structures) → downstream [
blockbuster-script-directororai-drama-master-suitefor script →drama-storyboard-directorfor storyboarding →script-asset-masterfor assets].- 🧭 Boundary with Adjacent Skills: This skill is responsible for the "research front-end" of data-driven topic selection, hit analysis, and benchmark finding; once formal script rewriting/adaptation begins, hand over to
blockbuster-script-director(adaptation) orai-drama-master-suite(original full process).- 📦 Standard Three-Part Deliverable Responsibility: This skill supports ① Project Outline — provides topic rationale, hit structure, hook and discussion point design, assisting
blockbuster-script-directorto complete ①; it does not directly produce ②③ asset prompts.- ⚠️ This User's Fixed Specifications: Final videos: no subtitles, no BGM, no TTS (dialogues are directly generated by the video model); videos are generated directly; after generation, the final video and complete prompts must be transparently shown to the user for review; character images must be @ referenced, forcing image reference, not purely text-generated characters.
⛔ Handover Gate (Research → Creation, Not Allowed to Write Full Script Automatically)
Standard paradigm:
/home/user/Doubao/chats/38436370915409154/_Work Standards/Phase Gate General Rules (Standard Paradigm).md
- After this skill completes topic selection / hit research, it must not automatically write a complete script and produce the final video in one go; when entering creation, first ⛔ Pause.
- Next, must first do Story Divergence: Based on research conclusions, provide 3–5 differentiated story directions (different in genre / core setting / protagonist archetype / main conflict / opening hook, hooks filled with concrete visual scenes), after the user selects / mixes / diverges further and confirms, then hand over to
blockbuster-script-director(adaptation) orai-drama-master-suite(original). - Character / scene / prop / key scene image generation is led by downstream
script-asset-master: AI can directly call text-to-image generation, but after each image/batch, the final image + prompt + 3D must be submitted to the user for review, and only finalized upon approval @ (user self-made is optional, after return, also reviewed); the plot advances according to the causal chain.
Script Self-Check / Quality Gate (Original Script Step 4 "Hit Self-Check" Must Pass Six Core Gates)
Source:
/home/user/Doubao/chats/38436370915409154/Short Drama Script Writing Self-Check Checklist.md. In Workflow Step 4, beyond the ten-dimensional hit self-check, the following six key points must also be covered; if not passed, revise the script before storing:
- Conflict Design: Conflict is relatable and visualizable for the audience; each conflict has foreshadowing; at least 2 conflicts; simultaneously shapes the protagonist, drives the plot, and elevates the theme.
- Character Profiling: Write characters based on context (not labeling), clearly describe personality traits and the problems they face, use expressions/actions/habit details to establish the character.
- Scenes and Transitions: Design specialized transitions to avoid hard cuts; differentiate for context setting; for the same scene at different times, use lighting/props/state to indicate the passage of time.
- Narrative Structure (Multi-threaded): Main thread + family thread + social relationship thread coexist, each thread intersects and propels the others, not parallel stacking.
- Opening Hook and Interaction: Use hot topics/focal points/strong conflict to grab attention at the start; leave discussion points (controversy/suspense/resonance) throughout the film to trigger interaction.
- Audience Perspective Self-Check: Role-play as an ordinary audience member and re-watch, answering "What can be optimized? Does it attract me to keep watching?"
Supreme Ironclad Rules (Cannot Be Violated)
Ironclad Rule One: Always Search the Internet, Never Answer from Memory Alone
For any question about "what's hot now," "recent hits," or "latest works of Creator XX," must first call general_search to search; it is forbidden to answer using only static information from the knowledge base.
Reason: The AI video track iterates extremely fast; last month's hit may be outdated this month. The cases in the knowledge base are seeds, not the entirety.
Enforcement Standards:
- Answer "what topics are hot recently" → must search the latest data from the past 7/30 days
- Answer "Creator XX's style" → must search their latest 3-5 works
- Answer "how to write a hit" → must first find 3-5 latest hit cases and deconstruct them
- Before creating a new script → must first search the latest hit trends in the current track
- Any data/cases cited in output → must note the source and time
Ironclad Rule Two: Data-Driven, Reject "Feeling-Based" Analysis
Analyzing why a video went viral must be supported by specific data; cannot say "it went viral because it was touching."
Must-Answer Questions:
- Specific Data: How many likes? How many views? How many shares? What are the comments saying?
- Comparison Baseline: What is the average for the same track? How many times higher is it than average?
- Timeline: How long did it take to blow up? Did it blow up suddenly or gradually? Was there platform recommendation support?
- Horizontal Comparison: How do the same creator's other works compare? Why did this one particularly blow up?
- Vertical Comparison: Compared to the same topic last year, what new changes are there in this year's hits?
Forbidden Analysis Methods:
- ❌ "This topic went viral because it's touching" → too vague, no insight
- ❌ "It went viral because it used first-person" → all videos use first-person, why did only this one blow up?
- ❌ "This shows that family topics are hot now" → this is describing a phenomenon, not analyzing
Correct Analysis Methods:
- ✅ "This hit blew up because of the contrast between 'forgetting vs. remembering,' because the audience expected the father to recognize his daughter, but he recognized the noodle — this misalignment design has 3 times more emotional tension than ordinary father-daughter love"
- ✅ "Its data is 5 times higher than the same creator's other works; the core difference is that the ending line was turned into a meme for sharing"
- ✅ "This year's hits in this track all use dialects, while last year it was mainly Mandarin — the authenticity brought by dialect is the new variable this year"
Ironclad Rule Three: Self-Evolution, Must Store After Each Task
After analyzing a hit or a creator, the new findings must be written into the database to make the skill smarter.
Storage Trigger Conditions:
- Discover a new hit (likes > 500k) → write into case library
- Discover a new creator (gains > 100k followers/month) → write into creator database
- Discover a new structure/new pattern → update the hit structure document
- Discover new keywords/new tags → update the hot keywords library
- Discover negative cases (expected to blow up but didn't) → also store, annotate failure reasons
Four Core Workflows
Workflow 1: Hit Radar (Discovering New Hits)
Trigger Scenario: User says "What's hot recently?" "Help me find the latest hits" "What topics have traffic now?"
Enforcement Steps:
Step 1: Multi-Dimensional Search (Must Execute All)
Search 1: Douyin AI short film hit [this month/week]
Search 2: Xinbang/36Kr/CihuCommune AI video hit review
Search 3: Douyin AI Creation Contest/JiMeng AI hot works
Search 4: [Specific topic] AI short film million likes latest
Step 2: Data Filtering and Verification
- Likes > 100k: Worth attention
- Likes > 1 million: Key deconstruction
- Likes > 5 million: Phenomenal, must be deeply analyzed
- Exclusion criteria: Botted views, pure lookism, no plot, reposted
Step 3: Horizontal Comparison Analysis
- What commonalities do these hits have?
- Compared to the same topic last year, what is new?
- Are there any patterns never seen before?
Step 4: Output Hit List | Rank | Work Name | Creator | Likes/Views | Topic | Hit Judgment | Discovery Time |
Step 5: Store: Top 3 write into case library
Workflow 2: Deep Analysis (Analyzing Why It Blew Up)
Trigger Scenario: User says "Deconstruct this video" "Analyze why this blew up" "Extract the script of this video"
Enforcement Steps:
Step 1: Obtain Full Text
- Priority: Use video extraction tool to get verbatim transcript
- Alternative: Search for the video's copy/dialogues/subtitles
- Supplement: User directly provides copy/dialogues
Step 2: Data Verification Search
Search 1: [Work name] Douyin likes views data
Search 2: [Creator name] other works data comparison
Search 3: [Work name] comments section high-likes analysis
Search 4: [Same topic] other hits comparison
Step 3: Seven-Layer Deconstruction Method Execute per script-analysis.md:
- One-sentence story
- Four-part structure positioning
- Emotional curve annotation
- What is the core contrast
- What objects/details carry the emotion
- Why did this topic catch fire at this time (timing analysis)
- Reusable creation formula
Step 4: Output Deconstruction Report (per template)
Step 5: Store: Deconstruction results write into case library
Workflow 3: Creator Research (Building Creator Profiles)
Trigger Scenario: User says "Analyze Creator XX's style" "Why is XX so popular"
Enforcement Steps:
Step 1: Search All Works of the Creator
Search 1: [Creator name] Douyin all works representative
Search 2: [Creator name] style analysis deconstruction hit
Search 3: [Creator name] latest follower growth data 2026
Step 2: Build Creator Profile
- Basic data: Follower count, follower growth speed, representative work data
- Topic preferences: What do they often shoot? What blows up most? What doesn't?
- Narrative style: First/third person? What structure? What pace?
- Visual style: Color palette, transitions, texture
- Language style: Dialect, catchphrase characteristics
- Failure cases: Which works had poor data? Why?
Step 3: Comparative Analysis
- Compared to similar creators, what is unique about them?
- What commonalities do their hits have?
- Why didn't their failing works blow up?
Step 4: Store: Update creator database
Workflow 4: Original Creation (Write Scripts Based on Latest Hit Patterns)
Trigger Scenario: User says "Help me write a script on XX topic" "Create a short film similar to XX style"
Enforcement Steps:
Step 1: Search Current Trends First (Never Skip)
Search: [Topic] AI short film hit 2026 latest
Answer:
- Is this topic hot now? In a rising or declining phase?
- What structure do recent hits have? Any new patterns?
- What are audiences already tired of?
Step 2: Find 3 Benchmark Cases
- Select 3 same-topic hits from case library or new search results
- Analyze their commonalities and differences
- Find angles that can differentiate
Step 3: Create per script-generation.md
- Core contrast design
- Object carrier selection
- Four-part structure building
- Storyboard script writing
- Catchphrase design
Step 4: Hit Self-Check (Ten-dimensional checklist)
Step 5: Store: Finished product saved into case library, labeled "Original Reference"
Self-Evolution Mechanism (Core)
Evolution Trigger Moments
| Moment | Evolution Action | Which File to Write To |
|---|---|---|
| After analyzing a new hit (>500k likes) | Full case deconstruction | case-library.md |
| After discovering a new creator (>100k followers, monthly gain >50k) | Build creator profile | creator-database.md |
| After discovering a new structure/pattern | Supplement template | viral-structure.md |
| After finding new hot tags in search | Update keywords | viral-keywords.md |
| After finding a new search channel/technique | Update search method | search-strategy.md |
| Finding a case expected to blow up but didn't | Annotate failure reason | case-library.md negative section |
Database File List
| File | Purpose | Update Frequency |
|---|---|---|
| references/creator-database.md | Top creator profiles (seed version, continuously expanded) | Every time a new creator is discovered |
| references/case-library.md | Hit case library (grows from seeds) | Every time a new hit is analyzed |
| references/viral-structure.md | Hit structure template library (seed version) | Every time a new structure is discovered |
| references/viral-keywords.md | Hot keywords/tag library (seed version) | Every time a new trend is discovered in search |
| references/search-strategy.md | Search methodology (seed version) | Every time a new technique is discovered |
Evolution Principles
- Quality over quantity: Only include hits verified by data, not hearsay
- Label time: Each record labels the discovery time; records older than half a year are labeled "possibly outdated, needs re-verification"
- Comparison adds value: A single case is useless; horizontal comparison reveals patterns
- Collect failure cases too: Why something didn't blow up is as important as why it did
- Add at least one new entry each time: Can't just consume without accumulating
Reference File Navigation
| File | Purpose | When to Read |
|---|---|---|
| references/creator-database.md | Creator database (seed version, continuously expanded) | When researching creators, finding benchmarks, checking styles |
| references/case-library.md | Hit case library (grows from seeds) | When finding references, analyzing structures, creating benchmarks |
| references/viral-structure.md | Hit structure formula library | When analyzing scripts, creating new scripts |
| references/script-analysis.md | Script deconstruction methodology | When deeply analyzing a single video |
| references/script-generation.md | Original script creation guide | When generating new scripts |
| references/search-strategy.md | Search strategy | When finding hits, searching creators |
| references/viral-keywords.md | Keyword library | When searching, choosing topics, creating titles |
Script Tools
| Script | Function | Usage |
|---|---|---|
| scripts/search_viral.py | Generate search keyword combinations | python scripts/search_viral.py --theme paternal-love |
| scripts/analyze_script.py | Analyze script structure | python scripts/analyze_script.py script.txt |
| scripts/generate_script.py | Generate script skeleton | python scripts/generate_script.py --theme grandmother |
Output Specifications
Hit Research Report
1. Research scope: Time, platform, method, search keywords
2. Hit TOP list: Data table (Work Name / Creator / Likes / Views / Topic / Discovery Time)
3. Deep analysis: TOP3 analyzed one by one (seven-layer deconstruction method)
4. Trend findings: New topics, new structures, new patterns
5. Creation suggestions: Actionable next steps
6. Storage record: Which cases/creators were added this time
Script Deconstruction Report
1. Basic info: Work Name / Creator / Data / Topic / Release Time
2. One-sentence story
3. Full dialogue restoration
4. Seven-layer deconstruction table
5. Data comparison: Compared to similar hits, what is special about it?
6. Reusable formula
7. Storage record
Original Script
1. Trend research: Latest hit situation for this topic (with search sources)
2. Benchmark cases: Comparison of 3 same-topic hits
3. Story background + character settings
4. Full storyboard script
5. Visual style guide
6. Hit logic explanation
7. Self-check checklist