One Unemployed Dev, Five Skills, and the Gating of AI Short Drama

The most honest README on GitHub right now might belong to a short-drama pipeline tool. Before it explains a single feature, eternityspring/shuohao-skills opens with a note from its author: he’s between jobs, this repo is a side exploration, and he’s open to remote work or a half-partnership arrangement of a few thousand yuan a month plus profit share. There’s a WeChat QR code and a paid community group for AI short-drama workflows. It is, in other words, a résumé piece with a business model bolted on — and it happens to be one of the more interestingly engineered things in the current wave of AI filmmaking tooling.

The attention is real. Directory listings show the project moving from roughly 3.2K stars to 6.1k–6.3k between snapshots, with one aggregator noting it’s actively maintained with a thin issue backlog. For a single-maintainer, Chinese-first repository of agent skills, that trajectory says something about where the demand is.

What it actually is

shuohao-skills is a collection of skills — structured instruction packages plus deterministic scripts — designed to be consumed by AI coding agents, explicitly Claude Code and Codex. You feed it a novel. Out comes the pre-production stack for an AI-generated vertical short drama: an adaptation outline, character design sets, art direction for scenes and props, episode scripts, and storyboards packaged for a video-generation pipeline.

The scope has grown faster than the documentation. The repository’s own README describes three skills — novel-characters, novel-outline, novel-art — while third-party listings describe five, adding novel-script and novel-storyboard. Either the README snapshot lags the code or the listings are ahead of what’s shipped; either way, the drift between the repo’s self-description and its catalog entries is a small tell about a project moving quickly with one pair of hands.

What the listings describe for the two newer skills is where the pipeline gets serious. novel-script produces episode scripts organized by scene and beat, with dialogue sheets aggregated per character and annotated with voice prompts — meaning the output is shaped to drop directly into text-to-speech. novel-storyboard segments the script into generation-length units with hard duration gates (segments capped at 15 seconds, individual shots at 2–5 seconds), cut points, model prompts, and an exportable production package. Per one Chinese-language review, the storyboard skill also ships built-in prompt alignment for MiniMax’s H3 video model — a sign the author is targeting a specific, current generation pipeline rather than abstract “AI video.”

The architecture idea worth stealing

The pipeline’s core design principle, as summarized by that review, is what it calls outline convergence: the outline, script, scenes, and characters iterate together, and the storyboard layer only renders — it makes no new decisions. That’s a genuinely disciplined stance in a field where the default failure mode is letting the model improvise at every stage and hoping continuity survives.

Anyone who has watched AI-generated serial content fall apart knows the failure: episode four’s protagonist has a different face, the pacing drifts, the prop that mattered in episode two evaporates. The commercial tools attack this with features — Boords sells a saved cast you reference by name, Storyboarder.ai claims market-leading character consistency, PicLumen promises recurring leads stay recognizable. shuohao-skills attacks it structurally: decisions converge upstream, and downstream stages are forbidden from introducing new ones. The storyboard step is a renderer, not an author. That’s the same philosophy that keeps rendering engines from rewriting your geometry.

The second idea is the quality gates, and this is the part I’d argue matters most. Each skill ships with scripted validation — the numbers vary by snapshot, but the counts run in the low-to-mid teens per skill (13 for outlines, 11 for art, and 10–17 for scripts and storyboards depending on which listing you trust and when it was taken). The critical detail is what they are not: they are not the model grading its own homework. They’re deterministic Node scripts checking structural invariants — durations within bounds, prompts matching shots word-for-word, required sections present. In a genre that runs on vibes, someone built a linting step.

The boring part is the good part

Under the hood, the engineering is almost aggressively unglamorous. The scripts use only the Node standard library — zero npm dependencies, nothing to install. Every skill carries a selftest that never calls a model and never spends quota, and the bundled examples double as test fixtures. There’s no CI, on the stated reasoning that the full selftest suite runs in about a second, so running it locally beats waiting on a pipeline. Installation is symlinks, so a git pull updates everything immediately.

This is the part that deserves respect. The entire AI-tooling ecosystem is currently drowning in prompt collections that are just markdown with confidence, and in npm packages that pull forty transitive dependencies to format a string. shuohao-skills is closer to the old Unix instinct: small, self-contained, testable units that a machine can verify without burning inference. The selftest-without-model-calls requirement is a hard rule in the repo’s conventions, and it’s the kind of constraint that seems fussy until you realize it means the deterministic layer can be trusted completely, separately from whatever the LLM did that day.

It also runs on your existing session quota — no separate API key — and image generation is optional, riding Codex CLI’s built-in image generation when present and skipping gracefully when not. As a product characteristic: this thing runs wherever your agent runs, at the marginal cost of tokens.

The market it’s undercutting

The context for the hype is a SaaS land grab. The pre-production layer of AI filmmaking has been colonized by subscription products in the past two years: Storyboarder.ai at $45/month claiming 250,000 creators, Boords at $26/month, Drawstory tiered from free to $150/month, PicLumen charging per second of generated video. One comparison piece relays invideo’s own claim of roughly $1,000 per finished episode for a three-person team — a figure that includes human labor, but still sketches the price of doing this commercially.

Against that, shuohao-skills is free, Apache-2.0, and local. The trade is obvious and not entirely favorable: the SaaS tools ship polished interfaces, client-review workflows, versioned animatics, and support. shuohao-skills ships markdown reports, JSON, and HTML you assemble yourself. If you’re an agency delivering storyboards to a brand client, Boords’ client-facing machinery is the rational purchase. If you’re a solo creator or small studio grinding out a 60-episode vertical drama season — which is, per the same comparison piece, a standard season length in this format — the economics point the other way.

There’s also a philosophical alignment worth noting. The ScreenWeaver comparison draws a line between all-in-one studios that get you to a demo fast and script-first pipelines that “survive episode 30.” shuohao-skills is unambiguously in the second camp: it produces structured, gated, machine-checkable intermediates rather than finished video, on the theory that the structure is what keeps a long serial from dissolving. The Storyflow rankings make a similar argument from the other direction — that short-form projects fail in pre-production, on the page, not in the edit. A tool that spends its entire effort on the page is betting on the same diagnosis.

The rough edges, honestly counted

The limits are visible and worth stating plainly. It has been verified only on macOS with Node 24; Linux and older Node versions are theoretically fine but untested, per the repo’s own admission. It’s optimized for short-drama narrative specifically — the MagicNetWorld review notes that ads and music videos could borrow the structure but would need to adapt it. Reports default to Chinese with English available, which shapes who finds it and how fast it travels. The star counts and gate counts disagree across third-party listings, which is less a scandal than a symptom: nobody has done an authoritative independent hands-on yet, and the MagicNetWorld review explicitly says it’s based on public materials without independent testing. One listing also notes a practical fragility — storyboard reports load images via relative paths, so files must stay at their prescribed directory level or the report breaks.

And then there’s the bus factor of one, with a twist: the maintainer is openly job-hunting. The paid WeChat group and the partnership solicitation in the README are, depending on your temperament, refreshingly honest or a signal about long-term stewardship. If the author lands a full-time role, does the pipeline keep pace with video-model churn — the thing that killed a hundred prompt libraries when their target models changed underneath them? The MiniMax H3 alignment is an asset today and a maintenance liability tomorrow.

Why it matters anyway

Strip away the short-drama specifics and what remains is a template: take a creative pipeline that everyone is rebuilding as a walled SaaS product, express it as agent-consumable skills, and enforce quality with deterministic scripts that cost nothing to run. The five-skill structure — outline, characters, art, script, storyboard — with convergence upstream and rendering downstream is a portable pattern. The selftest-without-model-calls convention is a portable pattern. The zero-dependency constraint is a portable pattern.

The commercial competitors will keep shipping nicer frames and better client comments. But the interesting bet here is that pre-production structure — the boring, checkable, convergent kind — belongs in open source, where a novelist with a Claude Code subscription and a weekend can find out whether their book survives episode 30. That the bet was placed by someone between jobs, with a QR code where a funding announcement would normally go, only makes it a more accurate portrait of how a lot of this industry’s actual work is getting done.