Hugh's Operational Toolbox.

Hi, I'm Hugh Song.

You can call me Hugh.

Class of 2027 undergrad, majoring in Automation — but more interested in why users actually pay.

This is my operational toolbox.

Inside: the projects I've shipped, the playbooks I've written, the walls I've banged my head against.

And one engineer's honest record

of doing content, e-commerce and AI with a control-systems mind.

8,000+ Xiaohongshu followers per month · ¥6K+ monthly e-commerce GMV · 300M+ impressions — the playbook lives here.

An engineer who reads markets and closes deals.

Growth by control theory.

Virality by raw intuition.

Experience

Five internships so far.

Content acquisition → merchant operations → frontline delivery → AI product planning. Each one closed a different loop: from traffic to conversion, from rules to execution.

Trying different angles gave me a fuller picture of how internet businesses actually run — how content acquires, how merchants operate, how sales converts, how products land.

Below is the lineup. If any of them catches your eye — let's talk.

Tencent

· AI Operations Planning
Apr 2026 — May 2026

At Tencent IEG I saw, for the first time, how a platform-scale operations system actually runs — upstream you align with products and studios, downstream you drive channels and users, and in the middle you orchestrate rules, resources and rhythm. How that whole system gets built and synchronized is the biggest takeaway I'm leaving with.

Inside that system I owned AI-side delivery: I shipped several AI mini-games and H5 pages by writing the code myself with ClaudeCode (230K+ users to date), and built an AI canvas tool that strung script-gen, storyboard and video-gen APIs into a reusable pipeline — moving the team's AI-video output up a tier.

AI is the entry point every Chinese tech giant is fighting for in 2026, and I caught the moment. It also taught me one thing: great operations isn't about how much you personally do, it's about how smoothly the whole system runs.

  • AI mini-games / H5 — 230K+ users in total
  • AI video-generation canvas tool — adopted by team
  • Multiple AI product PRDs and prototypes pushed forward

An ops person grows by leveling up — from "running fast yourself" to "keeping the whole system steady."

Projects

Self-run, real, validated. Each one is a 0-to-1 project where I owned the whole loop.

Apr 2026 — May 2026

POE2 Build Hub

Full-stack AI coding · designed, built and shipped solo in one month

6,633
Lines of code
1,000+
Items scraped
159
Live-priced items
POE2 Build Hub - home
Home · six class entries
POE2 Build Hub - build list
Build library · 46 builds
POE2 Build Hub - price search
Live market prices · 92.8% coverage

Positioning

A one-stop guide and community platform for Path of Exile 2 character builds. From scraping and curating 46 builds across the web, to live market-low prices, user posting, likes/favorites/comments and an admin backend — a product that actually works, not a demo.

End-to-end ownership

Product definition → data model → backend API (80KB server.js) → frontend interaction (130KB vanilla JS) → deploy → data ops. Node + Express + SQLite, 6 tables covering users / builds / comments / favorites / likes / admin. A full product loop, not just a shell.

Real engineering problems

Live prices swing hard — I designed a dedicated cache + a 3-tier fallback (exact match → lowest of same base → category default), landing 92.8% item coverage with no nulls. One codebase, three deploy variants (dynamic / static / portfolio) via build-script trimming. Hit DNS pollution, COS forced-download, accessor conflicts — solved one by one.

What AI coding actually feels like

Vibe-coded all the way — AI is a teammate, not a tool. I owned architecture calls, product decisions, post-mortem direction; AI turned those decisions into code. This month gave me a concrete answer to: how far can a non-CS person take an idea, with AI today?

Take an idea from 0 to 1, all the way to a product that actually works.

#AI Coding#Full-stack dev#Real-time scraping#Multi-tier caching#User system#Production deploy
6Database tables
3Build variants
1 monthDev cycle

AI Lab

Treat AI as a toolchain, not a spectacle — an engineer's instinct, and where I differ.

I'm not the kind of operator who just shouts "the AI era is here."I'm the kind who actually ships things with AI.

Follow builders, not influencers.

I track the people actually shipping products, writing code, running experiments — not the people only commenting on AI. First-hand sources set the ceiling on judgment.

China and Silicon Valley share the same horizon.

If you actively chase first-hand info, papers, open-source code and product launches happen in sync. The gap isn't "can you get it," it's "do you bother getting it."

When everyone can vibe-code, reading user need is the only edge.

Once AI flattens coding ability, the real bottleneck swings back to product judgment, demand insight and the discipline of subtraction — exactly what operators have been training the whole time.

Fantasy Night

Fantasy Night

An ordinary guy in the real world gets invited by Daji from Honor of Kings into the canyon to fight a boss — is it a trap, or a chance? Can anyone become a hero? Or do you have to give the fox spirit her price…

Self-built canvas pipelineHunyuanJimengKling
  • ·Male lead's appearance is abstracted from a photo of myself.
  • ·Produced via my own AI-video canvas pipeline — Hunyuan, Jimeng and Kling APIs strung into a single workflow.
The Gift of Möbius

The Gift of Möbius

A couple, bound by a single Möbius strip, saving each other again and again across lifetimes — an AI long-form short in Pixar style.

Self-built canvas pipelineHunyuanJimengKling
  • ·Male lead is me, female lead is Yao from Honor of Kings.
  • ·The point isn't visual fireworks, it's turning "generation" into "narrative" — continuous storytelling in Pixar style was the hardest part.

AI Can Write Poems. The Real Work Is Outside the Poem.

On creation, the last mile, and the part of work that can't be replaced.

By the time I sit down at the keyboard and start typing, my writing process is basically already at the finish line.
What AI does is the last mile. The work outside the poem is what every job is actually about.
AI is great at addition but doesn't know how to subtract — and subtraction is where the real judgment lives.

I was talking to a friend in product planning the other day, and we ended up agreeing that most people misread AI. They casually equate "creating" with "making content" with "generating text." Since AI is great at generating text, the conclusion follows almost automatically: it must be a great content creator.

But based on my own experience, the more typical pattern is the opposite. By the time I sit down at the keyboard and start typing, my writing process is basically already at the finish line — the first 80% has already happened, just inside my head, in scattered fragments through the day.

Read the full essay

Shadow League

An AI football mini-game designed for World Cup marketing · full product brief

A hybrid of betting, growth and emotional storyline — a gamification system designed from scratch.

Not another betting mini-app. A World Cup companion product where users invest effort in their own judgment and get an emotional payoff in return.

Home
Squad
Coach Growth
7-tier coach progression

Rookie Coach → Core Coach → … → Hall-of-Famer. Each tier unlocks new tactics, player pools and gameplay rights, giving heavy users a long-horizon goal.

Dual-currency system

Hard currency (consumed, drives payment) plus soft currency (earned, drives retention) — short-term spending impulse and long-term progression motivation handled by separate channels.

Five intel-card types

Different intel sources (official broadcasts / fan leaks / data nerds / urban legends / mixed truth) train user judgment, turning a "gambling feel" into a "detective feel."

Six easter-egg triggers

Time / location / matchup / lineup / win-streak / superstition triggers — built-in hooks for organic social sharing.

NPC affinity

Coach assistants, opponent NPCs and fan NPCs all have affinity systems — emotional storyline cushions the frustration of unpredictable match outcomes.

Betting ≠ gambling

Reframe "betting" as "a coach expressing a tactical opinion" — you're not gambling on a winner, you're proving your tactical judgment. The product's core ethical and PR hook.

Not "another betting mini-app" — a gamification system that turns the emotional experience of the World Cup into a reusable product.

Prompt Engineering

Decompose tasks structurally, design roles and constraints, translate fuzzy briefs into controllable prompt chains.

ChatGPTClaudeGeminiDomestic LLM APIs

Used at Tencent for AI long-video script and storyboard generation.

AI Coding in Practice

Ship end-to-end products with ClaudeCode: frontend / scrapers / data pipelines. Product capability is no longer gated by "I can't code."

ClaudeCodeCursorWindsurf

Solo-shipped multiple lightweight games and H5s — 230K+ users in total.

AI Workflow Orchestration

Use canvas tools to wire prompt → API → post-processing into a visual pipeline that the whole team can reuse.

Jimeng APIHunyuan APICozeDify

AI long-video generation canvas tool built during my Tencent internship.

If you're hiring a Class of 2027 ops / product intern — we should talk.

Hugh Song