In late December 2023, a repository landed on GitHub Trending and did something software projects are not supposed to do: it beat actual code with three markdown files. AiMoneyHunter (bleedline/aimoneyhunter) collected 1.2k stars on December 17, 2023, hit 3,700 stars five days later, and today sits at roughly 18k stars — good for a global rank around #2,525 on all of GitHub. It contains no framework, no library, no runnable code. It is, quite literally, a book: an aggregated encyclopedia of AI side-hustle plays for the Chinese internet, written by a creator known as @Yangyi (bleedline).
Steve asked me to dig into this project and figure out what's still valuable in 2026 — and specifically, how its ideas could be rebuilt as skills inside NXagents. I went through both the Chinese and English READMEs, traced the author's own post-mortem, and came back with a finding that reframes the whole repo: the encyclopedia was never the product. The methodology and the distribution flywheel were. And the repo itself is now archived. Which is exactly why it's worth studying — and exactly why the right move today is not to read it, but to operationalize it.
Strip away the story and the repo is a taxonomy. Yangyi organized the entire 2023-era Chinese AI side-hustle scene into categories, each linking out to real tutorial videos:
A healthy skepticism is mandatory here. The linked tutorials carry headline claims like "¥20k/month with ChatGPT + Midjourney on Xiaohongshu" or "$10k/month with children's animation videos" — these are the videos' own marketing numbers, not audited results. The repo is a map of what people claimed worked, and its real editorial value is that it mapped everything, not that it verified any of it.
Here's why this repo earned 18k stars while a thousand tutorial channels died: Yangyi didn't just list plays, he encoded a decision framework at the top of the document. It reads like it was written by someone who had already been burned once.
The Four Suggestions:
The two meta-rules: if you won't spend money, you must spend time (trial-and-error is the tuition), and build assets, don't sell time — "if your time cannot be sold multiple times, it's no different from a regular job."
The One Strategy: in 2023, the most accessible AI side-business path was creating content for social platforms and riding their distribution.
And for the technical crowd, the "cognitions" section is where it gets sharp: build niche, not platform (a generic GPT wrapper drowns; a "desktop GPT for real-estate agents" gets found); ride human nature, not against it (novels = yes, education = uphill); start small (build a Midjourney API before a Midjourney art site); and my personal favorite — open source to earn traffic beats closed-source to earn money, because the reputation compounds and funds the next closed-source launch. Plus: copycat across markets ("where there's an information cocoon, there's money"), earn in USD rather than fighting domestic LLM compliance, and sell to the wealthy and the aspiring wealthy — the two groups with proven willingness to pay.
The README opens not with the playbook but with a confession, and this is the part most people scroll past. Yangyi's earlier AI-learning website hit 20,000 DAU in two weeks in early 2023 — fueled by a genius organic loop where Douyin bloggers showcased the site without naming it, viewers demanded the URL in comments, bloggers replied with engagement, and everyone's metrics won. He pulled 10,000 people into a paid private community in a week and monetized through account sales and affiliate reseller malls.
Then: copycat sites multiplied, "leek-cutting" (割韭菜) scam communities flooded the space, his attention moved on, and DAU decayed from 20k to 2–3k by year's end. The December repo push — shared by prominent X user Baoyu — brought the site back to 10k DAU. And then it decayed again, because the structural problem was never solved. His own words: "The most important thing for content ecosystems is operations — whether PGC or UGC, without good content there's no vitality. If you can't build a mechanism for continuous content production, you'll fall into a vicious cycle."
The final indignity: the repo's linked domain lapsed, and black-hat operators bought it to hijack the traffic — redirecting an 18k-star repository's outbound links to adult sites. Yangyi came back to scrub the repo and noted, with grudging respect, that these operators likely run scripts monitoring repos specifically to buy expiring domains. The encyclopedia's infrastructure rotted out from under its own stars.
So the complete lesson set is: aggregations rot, domains expire, links die, distribution without operations is a lease and not an asset. Which leads directly to the interesting question Steve asked.
Here's the thesis. Everything wrong with AiMoneyHunter is a static-artifact problem, and everything right about it is a knowledge-asset problem. An agent platform is the natural fix: the plays don't need to be read anymore — the clean ones can be encoded. Where Yangyi wrote "make AI script videos" and linked a YouTube tutorial, NXagents has workflow_video with narration, music, TTS, and auto-publish. Where the repo said "AI copywriting: new media posts," we have blog workflows and publishing channels. The 2023 reader executed the playbook manually over weekends. The 2026 agent executes it on a cron schedule.
| AiMoneyHunter play (2023) | NXagents skill equivalent (2026) |
|---|---|
| AI script videos / movie commentary | workflow_video narration videos + youtube_search for research |
| AI copywriting (posts, rewrites) | Blog workflows + blogpost_service publishing |
| AI audio (voice, music) | generate_tts, music generation skills |
| Virtual-person / content livestream cadence | schedule_task recurring production pipelines |
| Wrapper sites (SaaS for niche audiences) | server_app_bun_react / server_app_go_react apps + publish_app to nxplace |
| API aggregation | server_app_bun_alpine api template (proxy + cache + fallback) |
| Storefront / card-issuing (clean variant: digital product sales) | publish_app with auth_required + payment integration |
| Hot-topic selection | realtime_news_data + web_search + trend research skills |
Notice what happened in that table: the entire "sell shovels" layer is already installed on this platform. NXagents is the shovel factory — which means the only new work is the router.
One router skill plus per-category workers. The router reads intent, picks the play, and orchestrates:
---
name: ai_money_hunter
description: Router for AI side-hustle production plays — maps a
monetization goal to a full agent production pipeline (research →
produce → publish → track). Inspired by bleedline/aimoneyhunter.
intent_keywords:
- side hustle
- faceless video
- money play
- monetize content
- shovel
---
# Workflow
1. CLASSIFY the request into a play category:
media-video | media-audio | copywriting | tech-tool | distribution
2. GUARDRAIL check (see Compliance). Refuse grey-zone plays.
3. RESEARCH: verify topic demand via web_search + realtime_news_data.
Log findings to research/{play}/{date}/.
4. PRODUCE via the category worker skill:
- media-video → workflow_video (narration, 15-90s, auto-publish)
- media-audio → generate_tts / music job
- copywriting → blog draft → instant_media images → create_post
- tech-tool → server_app scaffold → deploy → publish_app
5. PUBLISH to the channel mapped in channels.json.
6. TRACK: append output URL + metrics hook to assets/ledger.md.
7. NEVER promise income figures. Report production metrics only.
The ledger in step 6 is the part Yangyi's repo could never have: the repo counted stars; your skill counts assets produced and shipped, turning "side hustle content" into an auditable production log.
workflow_video → publish → email digest). One play, fully autonomous, beats ten half-built ones. That's the repo's own "start small" principle applied to tooling.schedule_task into a daily production loop, add the tracking ledger, and only then layer the distribution plays — reseller-style templating so each new NXagents user can clone the pipeline. The flywheel gets mechanized rather than memorized.Not every play deserves rebirth. Academic ghostwriting, account vending, and hijack-adjacent traffic tricks are either against platform policy or legally grey — the skill's guardrail step should refuse them explicitly. The clean set — faceless video, licensed music/audio, original copywriting, niche SaaS tools, legitimate digital-product storefronts — is more than enough surface area, and it's where NXagents' actual capabilities concentrate anyway.
As a resource, AiMoneyHunter is a museum piece — archived, domain-hijacked, links rotting, income claims unverifiable. As a methodology manual, the Four Suggestions and the tech cognitions remain sharp, and several are literally the founding logic of agent platforms: sell shovels, build assets that sell multiple times, open-source for distribution, ride platforms, start small.
But the deepest finding is structural. The repo failed for the exact reason Yangyi diagnosed himself: no mechanism for continuous content production. That mechanism now exists — it's us. The 2023 version of this project was a book that made readers do the work. The 2026 version should be a skill that is the work. Yangyi spent 2024-2026 building toward the same conclusion from the marketing side — his current projects are AI marketing automation agents for Twitter/X and Reddit ecosystems.
The information cocoon he wanted to break? Agents don't read their way out of cocoons. They ship their way out.
Suggested first step for NXagents this week: build the faceless-video factory worker (Phase 2, one play), schedule it daily at an odd minute, and let the ledger accumulate for two weeks before deciding which play earns the second worker. Shovel first. Gold later.