TopicPulse · Douyin Content Radar
Turning breakout content discovery into an executable topic pipeline for two creator IPs
AI Content Workflow Prototype / 05
Turning high-engagement Douyin content from smaller creators into a sustainable topic-production system through AI triage, human approval, deep analysis, and compliant rewriting for two expert IPs.
This is neither a trend chart nor an automated copying tool. Its goal is to reduce browsing, misjudgment, and file accumulation while preserving human decision authority, turning reference content into traceable, compliant topic assets ready for production.
Built for two creator IPs—one focused on digestive nutrition and one on traditional wellness—with separate content boundaries.
M0 delivers a responsive interactive prototype with Overview, Daily Discovery, Review Queue, Deep Analysis, Topic Library, and Asset Library.
Current candidates are demo data; real Douyin collection, persistent database, AI analysis jobs, and NAS writes are not yet connected.
Default candidate rules are no more than 30,000 followers, at least a 33% engagement ratio, and publication within 30 days; these are configurable filters, not claimed business results.
The Team Needs a Judgment System
The team relied on manual browsing to find reference content, with inconsistent breakout criteria. Videos were downloaded too early, creating local and NAS clutter; analyses were hard to reuse; and topics for the two expert IPs lacked a shared status system. Discovery, analysis, and production became disconnected.
The product narrows the problem to an operable loop: discover candidates consistently, download only approved items, and turn deep analysis into assigned topics with production states. It does not aim to cover all of Douyin; it serves the ongoing output of two creator IPs.
Decision Model: Discover, Approve, Analyze, Produce
The system first collects public metadata using keywords, seed accounts, and time windows, then evaluates follower limits, engagement ratio, topic fit, and an AI score. Operators do not download immediately; they first decide whether to queue, ignore, postpone, or approve analysis, then confirm the target IP.
Only approved items enter temporary download, transcription, keyframes, comment-need analysis, structural analysis, and compliance rewriting. After human review, a primary topic and alternatives move through topic pool, script, filming, published, and reviewed states, feeding the next round of discovery rules.
AI Recommends; People Decide
AI handles repetitive work: quantitative filtering, topic matching, rewrite-potential assessment, transcription and structure analysis, comment-need synthesis, and compliance warnings. It provides reasons and uncertainty but does not own download, assignment, or publishing decisions.
People decide whether an item is worth analyzing, which IP it belongs to, what must not be copied, and whether it can enter scripting and filming. Feedback never physically deletes source candidates; ignored items can be restored, and approvals and status changes are recorded.
A Download Gate and a Compliance Gate
Before approval, only thumbnails, public metadata, and source links are kept. MP4 files are downloaded temporarily only after approval, and only reusable, successfully analyzed assets are archived long term. Failed, ignored, or low-value files enter cleanup so the NAS does not become an unbounded dump.
Health content also requires a separate compliance gate. Treatment or cure promises, absolute effects, medical misinformation, testimonials, and privacy risks are flagged with safer alternatives. Reference content is for internal study and rewriting, not for republishing original visuals, audio, or captions.
PROCESS
Key Decisions
Prioritize breakout engagement, not just large accounts
The breakout ratio uses follower and engagement snapshots captured at collection time. Missing save counts are not treated as zero, avoiding false precision.
Download only after human approval
Separating worth reviewing from worth storing permanently reduces local and NAS clutter while keeping download accountability explicit.
Keep distinct boundaries for two creator IPs
AI can recommend an assignment, but operators confirm it; every topic keeps its source, IP, stage, owner, and deadline.
Make compliance rewriting part of the core flow
Compliance is not a final add-on; it is a required analysis output reviewed alongside the hook, structure, audience needs, and IP rewrite.
Role & Collaboration
I led product and experience: defining the breakout-content discovery objective, separating the two IP boundaries, and designing filters, human approval, asynchronous analysis, the topic state machine, asset retention, health-content compliance, and collection risk controls.
AI supported prototype engineering, interaction implementation, documentation, and workflow checks. I remained accountable for product judgment, rule priority, content compliance, evidence boundaries, and stage acceptance.
Validation, Outcome & Reflection
The confirmed stage is an M0 responsive interactive prototype with six modules, primary navigation, and candidate-decision flows. Its candidates and metrics remain demo data and must not be presented as real Douyin collection, operational AI analysis, or NAS writes.
The next stage starts with 20 real samples to validate field availability and filters, followed by 10 analysis reports to assess human usability. These are PRD acceptance targets, not outcomes already achieved.
The value of a content radar is not collecting more videos, but creating consistent selection criteria and preserving why an idea was worth producing as reusable knowledge.
A real loop exists only after collection, human approval, analysis review, publishing feedback, and retrospectives are connected. The prototype first validates decision ownership and workflow boundaries.
DEMO / 05
Interactive prototype demo
A 33-second walkthrough of overview, candidate review, deep analysis, dual-IP topic libraries, and asset management. The video demonstrates prototype interaction only, not a connected real-data pipeline.
