AIYOU

A node-based studio for producing AI short dramas, shipped as a React web app with a Tauri desktop shell. Twelve node types cover the pipeline — outline, episode split, character design (a three-view sheet plus a nine-grid expression sheet), storyboard script, 2K storyboard images, Sora prompt assembly, per-shot video, episode analysis — and you build a run by dragging nodes onto a canvas and connecting them. Generation goes through a provider layer with a per-category priority list and automatic demotion, so a quota error or a run of failures moves the work to the next model instead of ending it. The README states the cost of the sample episode as under two yuan.

Who built itThe repository is owned by this account, which has two commits in it. 131 of the 139 are attributed to a different account, superman32432432, under user@example.com — an address reserved for examples rather than a mailbox — and three to a third contributor. The README is written in the first person singular and reads as one person’s work — including the part where they grade it. The account that owns the repository has 12 public repositories and 8 followers.

1. 01

   The worklist pointed at a fork

   The gallery entry was lay950822/AIYOU: a fork, with no stars, one public repository on the account and a bio pointing at a Chinese domain. The upstream is yubowen123/AIYOU_open-ai-video-drama-generator — 136 stars and 49 forks, created 2026-02-04, last pushed 2026-03-05 — and both copies carry the same description, the same 110-versus-139 commits from the same authoring account, and the same example address. A secondary source that lists a fork as if it were the project is the failure this archive is built to catch: the record is filed against the origin, and the worklist entry is marked with where the file really lives. The fork also explains its own anomaly, which had looked like a GitHub bug — pushed_at reading earlier than created_at is what a fork looks like when the upstream history predates the copy.

2. 02

   The author grades it 30 out of 100

   Most READMEs sell. This one audits. Beneath the feature tables and the badges, the author writes that the project existed to train their own vibe-coding ability, that it is therefore incomplete and has plenty of small bugs, and then marks it: 30 out of 100 for the system, 50 for the prompts, on the grounds that the prompts are the part worth anything. The conclusion is stated as plainly — that it can serve as a tool for conversation and study, and not as a productivity tool — followed by the announcement of a rewrite intended to close the gaps that this one left open, and an open invitation for practitioners to argue with it. Two things about that are worth the archive’s attention: the score is attached to a repository with 136 stars and 49 forks, and it is the author’s own, which is the only kind of self-assessment this archive can quote without checking anything.

3. 03

   What six months of build artefacts weigh

   GitHub reports the repository at 126 MB. The source is a small fraction of that. Committed into the tree are npm-cache/_cacache, 628 files and 157 MB of npm’s own content-addressable cache — single blobs of eight megabytes, sitting under a folder that is named as what it is; a videos/ directory of twelve files and 71.8 MB of generated output; and a releases/ directory holding one packaged build of 15.8 MB. The same instinct shows up in the dependency metadata: the project has four lockfiles — package-lock.json and pnpm-lock.yaml at the root, and both of them again inside server/. None of it breaks anything, and all of it is what a repository looks like when committing is also the backup strategy.

4. 04

   The storage layer was designed five times

   docs/ holds nineteen files, and the ones about where generated media lives are: LOCAL_STORAGE_DESIGN.md (19 KB), LOCAL_STORAGE_PLAN_V2.md (41 KB), LOCAL_STORAGE_MIGRATION_PLAN.md (44 KB), LOCAL_STORAGE_IMPLEMENTATION_GUIDE.md (10 KB) and WEB_STORAGE_OPTIMIZATION_PLAN.md (29 KB) — a design, a second design, a migration between them, a guide to carrying it out, and an optimisation pass. The second design picks the File System Access API so a browser app can write images and video into a folder the user chooses, laid out as one directory per canvas and one subdirectory per node type, with a metadata index file at the root and a migration path for everything currently in localStorage. A separate 31 KB BACKEND_ARCHITECTURE.md answers the same question the other way, by moving the work to a local Express server with SQLite; its own opening line describes the change as replacing an architecture that called the AI APIs directly from the front end. Both documents are still in the tree, and the tree shows which one shipped: server/ exists, with a 80,707-byte index.js in plain JavaScript rather than the TypeScript the document had argued for, alongside db/ and three route files.

5. 05

   Model fallback treated as a feature rather than an error path

   The parts of this project that would be an afternoon’s work in most codebases have a guide of their own. MODEL_FALLBACK_GUIDE.md documents a priority list per media category — image, text, video, audio — that the user reorders in a settings tab, and beneath it an automatic demotion: quota exhausted, HTTP 429 or 503, or three consecutive failures, and the request moves to the next model down. Three failures marks a model unavailable and it is skipped; an hour later it is tried again. Health is shown per model as success rate, consecutive failures and the timestamp of the last error, the counters live in localStorage under one key, and every switch raises a notification that names both the model that ran out and the one that took over. It is a small piece of engineering with a large effect on what the product feels like: a canvas full of long-running generations that does not stop when one vendor’s quota does.

6. 06

   One person’s build log, as far as the record can see

   The history is 139 commits — 58 in January, 78 in February and 3 in March — of which 80 carry a model trailer: Claude Sonnet 4.5 seventy-four times, Opus 4.5 and Opus 4.6 three times each. The oldest commit is dated 2026-01-06, a month before the repository existed, so the work began before there was anywhere public to put it. 131 commits carry the example address user@example.com, which is why the contributor list reads as three accounts rather than the one the README sounds like. Two tags, v0.1.0 and v0.1.1, and no releases. The community left one pull request — a custom Gemini base URL, closed with no comment — and one issue, which is open and asks whether there is a video tutorial, adding that the author’s correspondent has no idea how to use it. The project also brought its own agent instructions: .claude/skills/aiyou-node-builder.md is 16,595 bytes on how to add a node to this codebase, and a vendored interface skill ships adapters for fourteen different coding tools. Nothing has been pushed since 2026-03-05.