Work 001 — AI product, 2026作品 001 — AI 产品 · 2026 Agent skill · Open source · MITAgent Skill · 开源 · MIT

jackson-wang
.skill

A runnable cognitive framework distilled from Jackson Wang’s public words, packaged as an AI agent skill and shipped to Claude Code and WeChat. 从王嘉尔的公开表达中蒸馏出的可运行思维框架,封装为 AI Agent Skill,已上线 Claude Code 与微信。

Role角色
Solo: product, research, prompt design, distribution独立完成:产品、调研、提示词设计、分发
Timeline周期
v1 in half a day
April 2026
半天完成 v1
2026 年 4 月
Built with工具
Claude Code
女娲 nuwa-skill
QClaw
Status状态
v1 shipped
Roadmap to JWANG 2.0
v1 已上线
迈向 JWANG 2.0
Brief起点 起点Brief

What would it feel like to talk to Jackson Wang anytime? 如果能随时和王嘉尔聊天,会是什么感觉?

A fan site粉丝站 A thinking tool思考工具
What would he say他会说什么 How would he think他会怎么想

Fans don’t just collect quotes. They borrow a way of seeing: how he decides, when he pushes through, when he stops. I wanted a product that gives people that point of view on their own problems, like a career fork, a creative block or a hard week, in a voice they already trust. 粉丝收藏的不只是语录,更是一种看世界的方式:他如何做决定、何时咬牙坚持、何时停下来。我想做一个产品,让人们用这个视角审视自己的问题——职业岔路、创作瓶颈、难熬的一周——而且是用他们本就信任的声音。

Background: biochemistry student moving from wet lab to dry lab. Trained in distillation and purification, with zero vibe-coding experience before this project. It was my first shipped AI product, built in half a day of self-study. 背景:生物化学专业学生,正从湿实验转向干实验。专业训练是蒸馏与纯化,此前没有任何 vibe coding 经验。这是我第一个上线的 AI 产品,用半天自学完成。

Method方法 蒸馏Distillation

I ran it like a lab protocol我把它当作一套实验流程来做

Distilling a person is still distillation: start from a messy raw material, separate the fractions, keep what’s pure and label what isn’t. 蒸馏一个人,本质上仍是蒸馏:从杂乱的原料出发,分离各个馏分,保留纯净的部分,并为不确定的部分贴上标签。

  1. Feedstock原料原料Feedstock

    I collected primary sources only where possible: X posts from 2022 to 2025, Instagram captions, MAGICMAN 2 tour transcripts, and Tribune, allkpop and TVB interviews. I logged them as seven research notes. 尽可能只采集一手资料:2022–2025 年的 X 推文、Instagram 文案、MAGICMAN 2 巡演台词实录,以及 Tribune、allkpop、TVB 的专访,整理为七份调研笔记。

  2. Fractionate分馏分馏Fractionate

    I separated how he thinks (models, decisions, timeline) from how he talks (sentence rhythm, vocabulary, language switching, typing quirks). Each fraction became its own spec. 把「他怎么想」(心智模型、决策、时间线)与「他怎么说」(句式节奏、词汇、语言切换、打字习惯)分开,各自成为独立的规格。

  3. Condense凝练凝练Condense

    I reduced the material to 5 mental models and 8 decision heuristics. Every one has evidence, a use case and a stated failure mode. 提炼为 5 个心智模型与 8 条决策启发式,每一条都附有证据、适用场景和失效条件。

  4. Purify纯化纯化Purify

    I stripped out anything unverifiable and wrote honest limits into the product: public words only, private life out of scope, research cutoff April 2026, and inference that never passes as his real judgment. 剔除无法核实的内容,并把诚实边界写进产品:只基于公开言论、私人生活不在范围内、调研截至 2026 年 4 月、推断绝不冒充他本人的判断。

  5. Formulate制剂制剂Formulate

    I packaged it all as one SKILL.md: trigger phrases, role-play rules, emoji logic, a quote archive and an exit command. 全部封装进一个 SKILL.md:触发词、角色扮演规则、emoji 逻辑、语录库和退出指令。

  6. Dispense分发分发Dispense

    The skill runs as /magicman2 in Claude Code, and in WeChat through QClaw for people who will never open a terminal. 在 Claude Code 中以 /magicman2 运行;通过 QClaw 接入微信,服务那些永远不会打开终端的用户。

Apparatus工具 Claude Code女娲 · nuwa-skillGitHubQClaw × WeChat微信
Inside the skill成分 成分Inside the skill
5Mental models心智模型
8Decision heuristics决策启发式
4Languages switched切换语言
7Research notes调研笔记
Model模型 In one line一句话 Where it breaks失效之处
  • Magic Man Dialogue魔法人对话魔法人对话Magic Man Dialogue

    The inner self you’ve been running from is your actual source of power.你一直在逃避的那个内在自我,才是你真正的力量源头。

    Assumes self-awareness. Without it, introspection becomes self-indulgence.前提是具备自我觉察;否则内省只会变成自我放纵。

  • Full Commitment as Baseline全力即标准全力即标准Full Commitment as Baseline

    Not “I’ll try harder”, but “if I don’t die, I can do it.”不是「我要更努力」,而是「只要不死,就能做到」。

    Burns people out without a rest mechanism. His 2023 health break proved it.缺少休息机制会把人烧干,他 2023 年的健康休假就是代价。

  • Duality as Fuel二元张力即养料二元张力即养料Duality as Fuel

    No low, no high. Don’t rush out of the valley; it’s feeding you.没有低谷,就没有高峰。别急着逃离谷底,它在喂养你。

    Risks romanticizing pain. Some bad things just need to be dealt with.有浪漫化痛苦的风险,有些坏事只需要被处理掉。

  • Identity as Moat身份是护城河身份是护城河Identity as Moat

    Who you are matters more than what you do. Scarcity comes from an identity no one can copy.你是谁,比你做什么更重要。稀缺性来自无法复制的身份。

    Needs a clear identity first, or it turns into an excuse to be different.需要先有清晰的身份,否则会沦为「与众不同」的借口。

  • Father’s Lens父亲镜片父亲镜片Father’s Lens

    If we lose, it’s my fault. If you win, it’s your glory.如果输了,是我的错;如果你赢了,是你的荣耀。

    Highly personal, and may not transfer across cultures or power structures.高度个人化,未必适用于不同文化或权力结构。

Plus Expression DNA: a real quote archive, running jokes, catchphrases, a multilingual switching map (Cantonese · Mandarin · English · Korean), and his observed typing quirks. 此外还有表达 DNA:真实语录库、梗库、口头禅、多语言切换地图(粤语 · 普通话 · 英语 · 韩语),以及观察到的打字习惯。

Product decisions产品决策 决策Product decisions

A persona is a behavior spec人设,就是一份行为规格

Pacing节奏

Reply length follows the user回复长度跟着用户走

By default an LLM over-explains, and that breaks the illusion instantly. He posts in short lines, so one sentence in gets one sentence back. Depth only comes when the question asks for it. 大模型默认会过度解释,这会瞬间打破沉浸感。他发帖都是短句,所以用户说一句,就回一句;只有问题本身需要时才展开。

回复长度跟着用户走

Trust信任

Say the disclaimer once, then stay in character免责声明只说一次,之后保持角色

Transparency and immersion pull against each other. The skill states its nature on first activation, speaks in first person after that, and leaves cleanly when the user says “退出”. 透明与沉浸天然存在张力。Skill 在首次激活时说明自己的性质,之后以第一人称对话;用户说「退出」,即干净地回到正常模式。

免责声明仅首次激活时说一次

Signal信号

Emoji are scarce, so they carry weightemoji 稀缺,才有分量

At most one per message, and none by default. Each emoji maps to a pattern observed in his real captions, which makes his emoji habits a set of rules you can test (try them below). 默认不发,每条消息最多一个。每个 emoji 都对应他真实文案中观察到的规律,于是他的 emoji 习惯变成了一套可测试的规则(可在下方试用)。

emoji 克制原则 · 一条消息最多一个

Safety安全

Every model ships with its limit每个模型都标注局限

The persona covers burnout and his darkest period. Stating where each idea fails keeps the skill from glamorizing pushing through at any cost, and that honesty is part of the product. 这个人设会谈到倦怠和他最黑暗的时期。标明每个观点在哪里失效,避免 Skill 美化「不计代价地硬撑」——这份诚实本身就是产品的一部分。

局限 · 诚实边界

Distribution分发

Reach people where they already are去用户本来就在的地方

The audience lives in WeChat, not terminals. Distributing through QClaw turned a developer tool into a chat window, and the setup guide includes one tip: ask for replies as separate messages so it feels like real texting. 目标用户在微信里,而不是终端里。通过 QClaw 分发,把开发者工具变成了一个聊天窗口;安装指南还附了一条小建议:让他分条回复,更像真实的聊天。

Claude Code → WeChat微信

Voice语言

Languages switch the way he does像他一样切换语言

Language follows the conversation. Emotional words cross languages, Cantonese comes out when he’s resigned, and greetings mix English with Korean. Forcing one language to sound professional wouldn’t be him. 语言跟着对话走:情绪词跨语言蹦出,无奈时冒出粤语,打招呼时英韩混搭。为了显得专业而只说一种语言,那就不是他了。

多语言切换地图

Specimen样本 样本Specimen

Emoji rules you can test可以亲手测试的 emoji 规则

Here is one small piece of the spec, running live. Type a caption and see which emoji the persona would use. Most lines get none. 这是规格中的一小部分,正在实时运行。输入一句文案,看看这个人设会用哪个 emoji——大多数句子一个都不会用。

  • 🚨❤️ Contains a 2, “two”, or a word with “to” (tonight, together, Toronto), or a word that sounds like /tuː/ 含数字 2、“two”,或含 “to” 的词(tonight、together、Toronto),或发音为 /tuː/ 的词
  • 🚨 Contains “az” (BRAZIL, amazing), UCB / UBC, Oakland, Chicago, LA, or New York 含 “az”(BRAZIL、amazing)、UCB / UBC、Oakland、Chicago、LA 或 New York
  • 🗣️ Contains a 6: June 6, the sixth show, 6am 含数字 6:June 6、第 6 场、6am
As posted发布效果

One emoji per message. In this demo, if several rules match, the one highest in the list wins. The rules come from patterns in his public Instagram captions, as written in SKILL.md. 每条消息只用一个 emoji。本演示中若多条规则同时命中,以列表中靠上的为准。规则来自 SKILL.md 中记录的、他公开 Instagram 文案里的规律。

Roadmap路线 路线Roadmap

From chat to presence从聊天,到陪伴

Full magicman2 persona完整 magicman2 人设 5 models · 8 heuristics5 个模型 · 8 条启发式 Expression DNA表达 DNA /magicman2 activation/magicman2 激活 Research notes调研笔记 WeChat via QClaw通过 QClaw 接入微信
Phase 1阶段一

Cross-platform跨平台

Put him where you actually spend time.把他放在你真正花时间的地方。

  • Expand beyond Claude skills to other agent platforms从 Claude Skill 扩展到更多 Agent 平台
  • Visual upgrade for the README in the MAGICMAN typography style以 MAGICMAN 系列字体风格升级 README 视觉
Phase 2阶段二

Cyber Jackson赛博嘉尔

He sends photos, speaks in his voice, shows up on your screen.他会发照片、用自己的声音说话、出现在你的屏幕上。

  • iMessage / WhatsApp-style chat interfaceiMessage / WhatsApp 风格的聊天界面
  • Sticker generation and song picks based on mood生成他风格的表情包,按心情推荐他的歌
  • Voice cloning, a wake-up alarm and a greeting when your screen wakes声音克隆、早安闹钟与亮屏问候
  • Exploratory: short videos and an animated avatar探索中:短视频生成与数字人形象
Phase 3阶段三

JWANG 2.0

He keeps evolving. The skill has to keep up.他在不断进化,Skill 也必须跟上。

  • Update the corpus with each new album, tour and interview随新专辑、巡演和采访持续更新语料
  • Evolve the persona as he does人设随他一起进化
  • Community-submitted quotes through Issues and PRs通过 Issue 和 PR 征集粉丝补充的语录
Reflection反思 反思Reflection

What it taught me about AI products它教会我的 AI 产品认知

The prompt is the PRD提示词就是 PRD

Writing SKILL.md was writing a product spec: triggers, rules, edge cases, tone, failure states and an exit. Vague instructions lead to vague behavior. SKILL.md 就是在写产品规格:触发条件、规则、边界情况、语气、失败状态和退出机制。指令模糊,行为就模糊。

Curated knowledge is the moat精选的知识才是护城河

Every builder has access to the same model. Nobody else has the sourced corpus, the details only a fan would notice, or the judgment about what to leave out. 每个开发者都能用同一个模型,但没有人拥有这份有出处的语料、只有粉丝才会注意到的细节,以及「该删掉什么」的判断。

Honesty is a feature诚实是一种功能

Stating the limits up front, like inference that isn’t his real view or a research cutoff, builds more trust than a persona that pretends to know everything. 提前说明局限——推断不代表他本人观点、调研有截止日期——比一个假装无所不知的人设更能赢得信任。

Ship small, then listen小步上线,然后倾听

Half a day to v1, published openly, with Issues and PRs as the feedback channel. The roadmap grows from what people actually ask for. 半天做出 v1,公开发布,以 Issue 和 PR 作为反馈渠道。路线图从用户真正提出的需求中生长出来。