Product definition
Structured video prompting, image prompting, and screenwriting as distinct workspaces that still flow into one another.
Project Overview
- Product name
- GENIUSHUANG
- Product category
- Full-stack AI Film Pre-production Workstation
- Current version
- v1.0
- Development
- May—Jul 2026, with ongoing maintenance and feature development
- My role
- Independent Builder
- Live product
- www.geniushuang.cn
Product boundary
The product interprets creative intent and scripts, organizes reference material, retrieves structured knowledge, and generates prompts and revised screenplays. It does not generate images or video directly, nor does it automatically invoke external video-generation models.
My Ownership
As the independent builder, I owned the complete path from problem definition and user-flow design to full-stack implementation, evaluation, and launch.
User experience
Designed idea input, text and image references, terminology rewriting, asset reuse, streaming output, iterative refinement, and bilingual UX.
Full-stack development
Built the interface, server endpoints, and model gateway with Next.js, React, TypeScript, and Route Handlers.
AI workflow
Implemented retrieval-intent routing, structured knowledge retrieval, a prompt compiler, and structured output protocols.
Knowledge engineering
Designed 16 codename knowledge databases containing 55,941 records, including 6,149 high-weight overlays.
Data & persistence
Implemented dual persistence across local JSON / file storage and Supabase Postgres / Storage.
Evaluation & reliability
Added input validation, connectivity tests, timeout and retry handling, model invocation, output-schema checks, and runtime evaluation.
Launch & iteration
Shipped the product, analyzed live usage, and refined its structure around user feedback, invocation anomalies, and failures.
Users & Problem
The target audience, their original way of working, and the core problems that made this product worth building.
Target users
Teams that publish continuously but lack a systematic cinematic prompting workflow
Built for AI creators, short-form video operators, e-commerce creators, MCN directors, and small brand-content teams—from a one-line idea to an executable film plan.
How users worked before
Start with a vague idea, then repeatedly add shots, style, characters, locations, and negative constraints across several chat windows.
Learn from or copy the structure of other creators’ long prompts, then manually rewrite it for each project.
Keep reference images, character definitions, and location rules scattered across local files and generation platforms.
When results fail, keep adding adjectives or restart—without knowing whether the real issue is framing, motion, continuity, model fit, or prompt structure.
Core pain points
Professional-language barrier
Users know they want a “cinematic” result, but may not be able to specify shot size, focal length, composition, camera movement, lighting, color, emotional rhythm, or action causality.
Cognitive and time cost
Long prompts demand extensive reading, decomposition, and trial and error—too slow for real production cadence.
Ambiguous intent
Short descriptions such as “premium” or “cinematic” allow many visual interpretations, so direct model input often produces generic results.
Continuity drift
When characters, locations, props, references, and versions are scattered, identity and visual rules drift across generations.
Unstructured correction
When a shot fails, users struggle to determine whether to change the subject, action, camera, lighting, style constraints, or negative prompt.
Problem statement
Users do not simply need “longer prompts.” They need a directing workflow that organizes professional judgment, supports low-cost revisions, and controls failure risk.
Core Workflow
From an initial idea to executable output, including product features, structured knowledge, and the decisions behind the system.
From idea to executable output
Add the idea and context
Enter an idea, script, storyboard, or task brief; upload references and reuse characters, locations, and props through named @assets.
Choose the task
Open the video-directing, image-prompting, or screenwriting workflow, then set language, duration, and creative direction.
Generate structurally
Infer retrieval intent, call relevant knowledge, and compile context; batch tasks can be decomposed into up to 10 items per run.
Review and revise
Review output by style, assets, shot content, and risk controls; cancel, edit, continue directing, or regenerate as needed.
Save and reuse
Recall previous sessions, save reusable assets, and carry them into future tasks.
Product feature walkthrough
A continuous walkthrough from idea input and structured output to screenwriting, asset memory, and external generation validation. Select a step below to switch the lightweight MP4 demo.
00:00 / 00:00
FLOW 01 / INPUT
Add an idea and references
Start with a one-line idea, then set language and duration and add reference material to establish context for structured generation.
MP4 · 5.6sProduct interface overview
SCREEN 01 / DIRECTOR
Director workstation
Add an idea, reference files, language, and duration, then enter the image- or video-prompt workflow.
Structured knowledge engineering
GENIUSHUANG is more than a fixed template. I turned years of film judgment, creative methods, reference cases, and high-quality outputs into a structured prompting knowledge system.
Skills / Templates
Easy to deploy, but often limited to fixed fields and generic wording. Flexibility and accuracy degrade across genres, styles, shot goals, and failure correction.
Structured knowledge + dynamic retrieval
Infer user intent first, retrieve knowledge related to the task, genre, style, and failure risk, then let the prompt compiler assemble a controlled context.
Experience & knowledge
Film-domain expertise and high-quality prompts from real production
Cleaning & deduplication
Normalize multi-source data and resolve duplicate content
Structured labeling
Unify fields, score quality, and complete missing content
Calibration & reinforcement
Use high-accuracy examples to strengthen other datasets
Intent retrieval
Select knowledge dynamically and compile a controlled context
1 + 15
output-framework core database + specialist databases
55,941
structured runtime records
6,149
high-priority calibration overlays
Key product decisions
Encode professional judgment in structured modules so users see an executable plan first, then revise it module by module.
Reference files, named assets, and history work together to preserve character, location, and prop rules.
SSE streaming, cancellation, continued direction, and editing reduce waiting and rework.
Handle one-line generation as well as long screenplays, shot lists, and batch tasks.
Technical Architecture
How data moves from user input through knowledge retrieval, the model gateway, structural validation, and persistence—and the stack behind it.
Technology stack
Next.js 16
React 19
TypeScript
Tailwind CSS 4
Route Handlers
Server-Sent Events
Request lifecycle control
OpenAI-compatible Chat Completions APIs
Configurable model gateway
Fallback
Supabase Postgres
Private Storage
Local JSON
Smoke Test
Schema validation
System architecture
Ideas, screenplays, reference files, and named assets.
Manages task state, attachments, history, assets, and the bilingual interface.
Handles input validation, rate limiting, file policies, and request lifecycle control.
The retrieval-intent module selects relevant knowledge; the prompt compiler assembles controlled context and structural protocols.
OpenAI-compatible Chat Completions with connectivity testing, timeouts, cancellation, retries, and model fallback.
Streams results to the frontend and verifies that critical modules and batch items are complete.
Local JSON / file storage supports development and offline fallback; Supabase supports production deployment.
Verified Outcomes
Live usage, iteration driven by user signals, engineering evidence, and the methodology behind every metric.
Live usage outcomes
Iteration driven by user signals
24.1% of deduplicated needs explicitly mentioned references or consistency, validating @asset references, reference files, and history reuse as core-path features.
Revision / redo requests and negative constraints each represented 17.2%, supporting edit, cancel, regenerate, and risk-control modules.
Historical batch retry failures led to batch detection, itemized output, schema validation, and repair retries.
Image prompting represented 52.5% of custom interactions, making it the clearest high-frequency entry point.
Engineering & evaluation evidence
Measurement notes
Anonymized access identifiers are server-side hashes of IP addresses, not
unique people or accounts. The 47.1% figure indicates repeat interaction
within the observation window, not retention. The 4.92/5 score is a
lightweight, fixed, rule-based evaluation of fixed outputs—not a user rating,
third-party score, or live model-quality metric.
GENIUSHUANG
From problem definition and knowledge engineering to full-stack delivery and live-usage validation, this is a 0→1 AI product I built and continue to maintain independently.