Tony | AI Engineer

AI engineer in Melbourne. I build AI agents that run real work, then show you exactly how. Copy-paste AI workflows daily. Also building GulaStories.

[Tony | AI Engineer @tonyaibuilds]

I gave Claude one prompt: make an animated Halloween pumpkin town that lights up one house at a time at night.

32 seconds later: 16 pumpkin houses grown, carved and lit, bats, a ghost and a soundtrack, all written as code. No stock assets.

Happy Halloween from Melbourne.

[Tony | AI Engineer @tonyaibuilds]

@sundyme 绝大多数传统公司，制造业公司，都还在阶段1摸索

[Tony | AI Engineer @tonyaibuilds]

@binghott I stopped searching sessions. Every agent writes a short report file when it finishes, and grepping those is way faster than hunting for the chat it happened in.

[Tony | AI Engineer @tonyaibuilds]

@brookejlacey Same thing on a much smaller scale for me. The agents can draft and post all day, but anything that talks to a real person still waits for me, and that queue is the slow part. Did she say which people stuff eats the most time?

[Tony | AI Engineer @tonyaibuilds]

@omkarships The one thing I can't do from the app: launchd kicks off claude -p on a schedule every morning, so the agents run without me opening anything. Headless runs and piping the output into other scripts is where the terminal wins for me.

[herdr @herdrdev]

*ehm ehm* BIG ANNOUNCEMENTS DAY 📣

two new things cooking in herdr labs 🐏🧪
a brand new herdr-core, and a desktop app built on it. open source.

first, the app: software that fits you. every part of it is a file your agent can change. you just ask.

[Tony | AI Engineer @tonyaibuilds]

@CyanCloud77 确实宝藏，关注一下

[Tony | AI Engineer @tonyaibuilds]

@yetone 重点，UI可以 完全自己调。非常舒服。

[Tony | AI Engineer @tonyaibuilds]

@yetone 我的做法：三件套。herdr 管多台 Mac 上的 Claude Code 会话（开工作区、派指令、看输出、采纳结果），otty 把终端变成手机能用的网页控制台，tailscale 把家里/公司的电脑和手机连成一个私有网络，走到哪都能连。躺床上也能派活、盯进度。

[美酱AI @meijiangAI]

月 30 万从「单价 × 销量」反推。

最后这步。

从月收 30 万反推设计。

500 元 × 600 份 ＝ 30 万元
1,000 元 × 300 份 ＝ 30 万元
3,000 元 × 100 份 ＝ 30 万元
5,000 元 × 60 份 ＝ 30 万元

同样 30 万，价格不同所需销量完全不一样。

这块也问 ChatGPT↓↓

「这篇 note 如果定价 500 元、1,000 元、3,000 元、5,000 元，

各自需要的
・月销量（达成 30 万）
・对应的信息量
・读者期待的价值
・购买难度
・适合的推广策略

分别比较一下。

根据我的成绩、一手信息和内容质量，理由充分地建议最合理的价格。」

总结一下，

选赛道
→锁定读者
→写 note
→X 引流
→数据迭代
→价格优化

这个顺序。

ChatGPT 再能写优美文案，
没人想要的 note 也卖不出。

要冲月 30 万，
比起写文章，先要设计「卖给谁、什么产品、什么价格」。

这才是最关键的。

[elvis @omarsar0]

Learn to build evals, folks!

You're not really doing recursive self-improvement (RSI) if you are not doing evals. And don't tell me your agents are generating the evals for you. They are pretty bad at it.

I have found it really interesting and useful to give agents a concrete way to measure results and stay on track with a long-term goal. It's easier to hill climb when things are clearly measured and well defined. Probably one of the most important skills anyone in the space can have today.

It doesn't matter how many agents and intelligence you throw at a problem; you are not going too far without proper evals.

[Tony | AI Engineer @tonyaibuilds]

Anthropic donated Agent Skills to the Agentic AI Foundation (Linux Foundation).

A skill is a folder with a SKILL.md file that teaches an agent one task. Anthropic says 70+ platforms support it, including ChatGPT and Cursor.

Source: Anthropic blog, Oct 22

[Tony | AI Engineer @tonyaibuilds]

6 one-line prompts that make AI check its own work:

1. List what you'd need to know first
2. List 3 ways your answer could be wrong
3. Rate confidence 1 to 10
4. Review it as a strict reviewer
5. Show the steps
6. Say "not sure" instead of guessing

Full wording in the images.

[山羊赚钱笔记 @shanyanggm]

小红书全套运营 Skill🫨🫨🫨🫨🫨🫨🫨🫨🫨🫨🫨🫨

1 | 选题策略
Skill 名称：content-strategist
功能：把平台规则翻成可执行清单，帮助你理顺选题方向与内容优先级。
适合人群：不会规划内容、想做账号定位的人。
核心亮点：关键词前置、标题长度控制、A/B 备选思路。
GitHub：https://t.co/mDax23SXtk

2 | 数据诊断
Skill 名称：dbskill
功能：给选题和标题做体检，分析哪里不吸引人、哪里有 AI 味。
适合人群：标题没感觉、点击率不稳定的人。
核心亮点：诊断清单、爆款标题公式、直观定位问题。
GitHub：https://t.co/1GxhHyp3Zn

3 | 文案生成
Skill 名称：xhs-writer-skill
功能：从卖点分析到正文、标题、标签，输出完整小红书笔记。
适合人群：想快速出稿、不想从零开写的人。
核心亮点：爆款拆解、卖点抽取、可直接发的成稿结构。
GitHub：https://t.co/6RX4LJPQEQ

4 | 热点联动
Skill 名称：writer + trending
功能：把写作能力与热点捕捉结合，让日更更容易蹭上热度。
适合人群：更新频率高、盯热搜词、需要日更灵感的人。
核心亮点：热点切入 + 爆款结构，少在工具间来回切。
GitHub：https://t.co/xBwjDdsgwB

5 | 封面配图
Skill 名称：space-image-studio
功能：通过一句话生成小红书封面或配图，快速制作风格化图片。
适合人群：不擅长设计、想快速出图的人。
核心亮点：4 类类型 × 12 种风格，适合封面和配图实验。
GitHub：https://t.co/ZCoHVmZAq0

6 | 图文卡片
Skill 名称：guizang-social-card-skill
功能：把文章、截图、笔记整理成套图文卡片与封面。
适合人群：需要制作组图、教程图、知识卡片的人。
核心亮点：Editorial / Swiss 两套视觉系统，版式更成熟。
GitHub：https://t.co/yyEHbh6eg7

7 | 排版分发
Skill 名称：html-anything
功能：把 Markdown、表格、笔记转换成单文件 HTML，方便多平台发布。
适合人群：长文作者，以及需要在公众号、知乎、小红书同步分发的人。
核心亮点：无需 API Key、快速生成页面、排版友好。
GitHub：https://t.co/9yk6rLJNdP

8 | 全链路工作流
Skill 名称：xhs-ai-workflow
功能：从调研起草、配图到发布，串起完整小红书流程。
适合人群：想优化工作流程、实现内容生产系统化的人。
核心亮点：完整的 AI 协作流程，适合长期内容运营。
GitHub：https://t.co/L3ekgRtmXV

[Tony | AI Engineer @tonyaibuilds]

What does a forward deployed engineer (FDE) do?

They go to the customer. Picture a bakery: sit with the owner, learn how orders, suppliers and rosters really work, then wire AI into that.

The model is the easy part. The plumbing and trust are the work.

Bakery is illustrative.

[Tony | AI Engineer @tonyaibuilds]

@mardehaym When one missed field means someone doesn't get paid, I'd keep a human click before anything gets submitted, at least until the error log has been boring for a few months. What ended up convincing the CTO, accuracy numbers or the review step?

[Tony | AI Engineer @tonyaibuilds]

@IamAroke I think the skill moves more than it disappears. Spotting what's subtly wrong in a diff is still problem solving, and it's hard if you never wrote the loop yourself. Having juniors write it first, then compare with the model's version, seems fair.

[Tony | AI Engineer @tonyaibuilds]

@itz_komal01 Part of it is the feedback loop. A broken web app shows you in seconds, but a database bug can sit there quietly for weeks. On the complex stuff I'd want the tests written before any vibing starts.

[Tony | AI Engineer @tonyaibuilds]

@snowmaker Threads are the part worth stealing. My agents each write a report file and one coordinator reads them, and the hard bit turned out to be deciding which message needs a human. How do you decide what pings you?

[Tony | AI Engineer @tonyaibuilds]

9 AI terms every manager should know, one line each:

Model, prompt, context window, token, hallucination, agent, RAG, MCP, evals.

The one people skip is evals: tests that check the AI's answers before and after a change. Ask any vendor how they run them.