Story

Hi fellow HNers,

I am guld and I created ROE.md (https://github.com/guld/ROE.md), a proof of concept single Markdown file that let's you raise your own custom personal OpenClaw-like AI assistant from scratch.

With a very stripped-down feature-set of course. (For now. Still alpha)

All you need is a little vibe code magic from your favorite AI coding tool. Select your favorite programming language (yes really) and hope that it has been grokked by the SOTA models. (I am looking at you Ruby programmers).

I came up with the idea last week when I first heard of the OpenClaw craze. Listening to steipete's wild stories I was wondering... if the agent can create all these skills by itself... what if we spin it further and let the agent create itself from a single Markdown file. This should work!

How to use it?

Your coding agent of choice should load the ROE.md markdown file, you answer a few questions and it builds a minimal OpenClaw-like personal assistant for you. It comes with a single build in tool (execute_bash).

This is dangerous I know, but it works.

Anyways... I tested it with OpenCode and the Kimi-2.5 model and Python selected. (OpenCode is great!) With a few questions answered and some bugs fixed the custom agent works nicely in CLI mode with a locally running gpt-oss-20b through LM Studio.

How did I create the ROE.md file?

Mostly by hand, with a very high level overview of the architecture, almost no architecture at all. I rather focused more on adding a lot of example call-responses and used first person singular to improve prompting (Thanks to a comment thread on here where postalcoder verified that first person makes the difference with Gemini)

The next few days I want to extend the ROE.md file to contain more succinct API examples to make sure it works with more languages.

- TODOs alot... Fix, Test, Repeat

Challenges:

- Getting more programming languages to work in as few prompts as possible with the big local models such as Kimi-2.5. Tested Bash, Python, Go, Ruby, and a few others but Python works best.

- Not getting to specific in pseudocode.

- Keeping its structure general enough to be applicable to most programming languages.

- Making it work with weaker local models.

For more details check out the mini-tutorial in the README (https://github.com/guld/ROE.md).

This project and the power of vibe coding brought back the enjoyment I felt when I started out writing my first line of code on an "empty piece of Windows Editor paper" more than two decades ago.

What do you think? Is vibe coding crazy cool or is it just me?