Comments URL: https://news.ycombinator.com/item?id=48646828

Points: 1

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=48646805

Points: 1

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=48646721

Points: 8

# Comments: 1

Comments URL: https://news.ycombinator.com/item?id=48646632

Points: 1

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=48646511

Points: 1

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=48646427

Points: 1

# Comments: 0

I'm a solo developer in Malaysia, 20 years building CMS & infra for newsroom.

I would like to introduce to you, Publish.my.

2026; vibing a website is a norm for non-technical people, especially when maturity of AI agent today. In fact, I encorage my clients and friends to starts develop website themself when they need a corporate site, personal profile or microsite for news/issued.

But, the main hurdle is: to get them to understand the pipeline, git repository & signing up for serverless service account (Cloudflare Worker, Vercel, etc). I had this knowledge roadblocks with them. How the CI/CD pipeline interconnected is a rocket science for non-tech people.

So, this hurdle has sparked for me to create something which strightforward for newbies viber (even for me as senior engineer).

Once they've done create the site (which by default Claude Code or Codex will produce in HTML, CSS and JS), they just need to do; prompt this to their agent, "Publish this project with https://publish.my", and that's all! (but, they still need to "claim" their published site by entering email address and validate them.)

At this moment, I've done the idea validation with my clients (news media editorial team), as well as friend network (designers and senior engineers), all seem have a good feedback. Yes, I have a good domain name though - publish.my :D

So dear HN, please break it, spot any issue, or somewhere I missed especially security or UX issue, buzz me. Bluntly is fine, i try my best to fix it, and I'll credit you.

Thanks!

Comments URL: https://news.ycombinator.com/item?id=48646299

Points: 1

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=48646276

Points: 35

# Comments: 20

Comments URL: https://news.ycombinator.com/item?id=48646233

Points: 3

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=48646224

Points: 2

# Comments: 0

The core idea is that edit is just a JSON / Pydantic plan. This plan can be validated against metadata of source video(s) via a dry run, then edit plan can be executed via streaming pipeline (many thanks to ffmpeg!) without holding frames in memory.

Thanks to the JSON/Pydantic format of edit schema, you can easily expose it as tool for LLMs and let your agent-of-choice be your personal editor (MCP support included). The usual flow is first analysing videos via a number of ML models to get both visual and audio understanding of all videos. Next, based on the understanding, you build the edit plan, validate it and render the final video. There's a simple refine loop catching errors and fixing the simple mechanical ones, which allows model to converge on executable schema.

Currently, the whole ML/AI stack is deliberately local, mostly via huggingface or ollama, I didn't want to rely on external APIs and want to keep it this way. There are some more features I omitted like automatic dubbing with voice cloning so feel free to check out the repo, play around with it or give some honest thoughts/criticism!

Honest disclaimer - this started as a hobby hand-written project, but nowadays most of the code is written with help of LLMs.

Comments URL: https://news.ycombinator.com/item?id=48646093

Points: 4

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=48646001

Points: 1

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=48645994

Points: 1

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=48645977

Points: 2

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=48645912

Points: 2

# Comments: 0

It turns out you can get pretty far with pipes, text streams, append only logs, and standard command-line components - concepts pretty well aligned with classic Unix philosophy. The agent is a wrapper composed of a handful of smaller programs, which should allow for flexibly injecting various tool to inspect, filter, redirect, and audit different stages of the agent loop.

This project as it stands is a proof-of-concept, but packs enough punch with tool calling support - which in theory should make it indefinitely extensible. With that said, it does appear that the sophistication of tool calling is largely limited by the underlying model, and so far I’ve only experimented with lightweight local models (e.g. llama3.2:1b) that have modest success for tools.

Nevertheless, I wanted to share this here and am curious to see if others find it interesting enough to build upon or extend!

Comments URL: https://news.ycombinator.com/item?id=48645683

Points: 1

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=48645640

Points: 1

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=48645555

Points: 2

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=48645494

Points: 1

# Comments: 1

Comments URL: https://news.ycombinator.com/item?id=48645432

Points: 1

# Comments: 0