Business judgment on AI products
Short-video writers search a local script library by plot tags and instantly find matching references, instead of scrolling feeds and jotting notes.
01
Start inside the user's day · Public facts + observable behavior · 2026-09-01
Short-video writers need to create original Douyin and TikTok story scripts and seek quick references.
Currently writers may manually scroll platforms or use note tools, but lack structured search.
Finding reference scripts requires scrolling feeds and taking notes, which is time-consuming and inefficient.
The trend is that short-video output is no longer the bottleneck — what to write and what to copy from is. Don't sell script generation. Build searchable reference libraries for story creators and short-drama writers; give away the engine and license the corpus.
Its public repository has 354 stars and 8 forks, showing developer attention; repeat use and payment are not yet verified.
1. Where the database came from and how it's licensed — this single question decides how long it lives, above every other metric; 2. Whether the database keeps growing — a reference library that stops updating is worthless within six months; 3. Whether a payment path appears — "separately licensed" …
Worth dissecting. Its public repository has 354 stars and 8 forks, showing developer attention; repeat use and payment are not yet verified.
The trend is that short-video output is no longer the bottleneck — what to write and what to copy from is. Don't sell script generation. Build searchable reference libraries for story creators and short-drama writers; give away the engine and license the corpus.
Code is MIT; the database is licensed separately. The README is explicit:; this repo does not contain the real database, and the database repo declares its own license. 1. Where the database came from and how it's licensed — this single question decides how long it lives, above every other metric; 2. Whether the database keeps growing — a reference library that stops updating is worthless within six months; 3. Whether a payment path appears — "separately licensed" …
Short-video writers search a local script library by plot tags and instantly find matching references, instead of scrolling feeds and jotting notes.
Its public repository has 354 stars and 8 forks, showing developer attention; repeat use and payment are not yet verified.
An English validation note will follow from the public evidence.
The product claims to help users complete: “Short-video writers search a local script library by plot tags and instantly find matching reference”. User evidence has not yet verified pain intensity or the cost of doing without it.
The assessment is recorded; an English explanation is pending.
The assessment is recorded; an English explanation is pending.
The assessment is recorded; an English explanation is pending.
02
Market comparison
The Chinese–English market comparison is not complete yet. A conclusion follows only after its coverage and verifiable evidence are recorded.
03
The call and next move come first; the full read retains the evidence and counterevidence.
A short-video script library you search by story tags, installed inside an AI agent as a source of reference material.
The GitHub account liujunxibaba, a Chinese developer.
The engineering details give away his actual setup: the installer is PowerShell (a Windows user),
it installs under $HOME\.codex\skills\ (Codex, not Claude Code), and nothing touches the network.
This is someone using it to do his own work, not a demo built to be looked at.
search "campus teacher-student misunderstanding conflict-in-first-3-seconds" --top-k 5,
pulling matching script samples by tag combinationWhat it deliberately does not do: it doesn't write scripts, make video, or handle shooting. It only does the step of finding the right reference sample.
A short-video writer looking for references actually does this: open the app, scroll comparable content, manually note the good ones, build up a personal document, and then fail to find anything in it next time.
The old behavior is unmistakable, and it's pure manual labor. A writer's time should go into adapting, not into "I remember something like this but I can't find it."
That search example says a lot about how well he understands the work: "campus / teacher-student misunderstanding / conflict in the first three seconds" — the first two are subject matter, the third is structure. Anyone who indexes on "conflict in the first three seconds" has written short-video scripts with their own hands.
Code is MIT; the database is licensed separately. The README is explicit: this repo does not contain the real database, and the database repo declares its own license.
The split is deliberate: give away the engine, sell the content. Anyone can read the code, but it doesn't run without the database, and the database terms are elsewhere.
Read: the pricing logic holds and it's healthier than pure open source. The risk is provenance — if the script samples were scraped from a platform, its copyright position is as fragile as open-kimi-ppt-skill's was.
| Dimension | Read |
|---|---|
| Founder-product fit | High. "Conflict in the first three seconds" is not a dimension an outsider invents |
| Product insight | High. It attacks "find the reference" rather than "write the script," dodging the most crowded step |
| Execution quality | Middling. Local search plus a health check is complete work, but not technically hard |
| Timing | Good. Short drama and creator-led storytelling are hot, and every tool is bunched at the generation end |
This project and shuohao-skills are two executions of the same judgment: both sidestep generation and work on the step before it.
shuohao turns a novel into shoot-ready material; this one finds the right reference sample. Both bet on the same thing: the bottleneck in AI short video is input, not output. The models can already write and shoot. What's stuck is deciding what to write and what to model it on.
The transferable rule: when generative capability is in surplus, the scarce things are judgment and a material library — the first is hard to productize, the second isn't. Which makes building the library the most concrete business available at this stage.
The cost comes in two layers. Copyright first — if the database was scraped, it can repeat open-kimi's ending at any time, and he clearly senses it (splitting the database out with its own license is risk isolation). Then the moat — search itself has no technical barrier, so the moat equals the quality and size of the database, and that requires continuous manual curation that can't be automated.
The Windows + PowerShell + Codex combination is worth noting too: that isn't the mainstream developer stack, which says his target user is someone making content on Windows, not a programmer. That judgment may well be right.
Index reference material by structure, not by subject. Anyone can tag a topic; "conflict in the first three seconds" is what a writer is actually searching for. Choosing structural dimensions over subject dimensions is what makes a reference library worth opening twice.
Business model: open the code, license the data separately. That's a healthier structure than a free tool plus a paid chat group — a group has to be maintained by a person, a database accumulates as an asset. If you want to productize a method, this is the shape worth copying.
Risk note: read this next to the other case from the same day (a 1,600-star repo wiped for copyright). Provenance is a live risk for anything built on a material library. If you build one, the sources have to be clean from day one.
Worth watching. The product judgment is right and the structure is sound, but everything hangs on whether the database is legally clean, and public information doesn't say. Come back in three months; whether it's still there is the answer.
05
Use these searches when the official site is missing or the current link is only a lead.