Introduce "Responsible Vibe Coding Recommendations" for LLS contributions.

[rhuss]

There has been a recent surge in PRs where unreflected AI code-assistant usage came into the way.
@leseb, @francisco, and I thought about this a bit and came up with a proposal for some responsible AI coding guidelines.
We call it the "Responsible Vibe Coding Guide" :-), which are some common sense rules that help in this new world of AI vibe coding.
It contains a lot of common sense that is also valid for non-AI usage, but AI-usage amplifies the risk of many monster PR that change also unrelated stuff.

You can find the guide here for now --> https://vibe-coding-manifesto.github.io (Source: https://github.com/vibe-coding-manifesto/vibe-coding-manifesto.github.io/blob/main/AI_MANIFESTO.md)

It's still a work-in-progress (i.e., we will rename and move away from "manifesto," which is already somewhat overused and a bit too prescriptive, so the URL will change, too).

We propose to add this file, maybe top-level, to the LLS repo (so Cursor and co. can find it, too), but before that, it might be cool if we could adapt this to the specific needs of the LLS community.

If you think this is a good idea, I'll create a PR for that file, and we can take it from there.

cc. @ashwin

[ashwinb]

Thank you for driving this. At least as far as llama-stack is concerned, what do you think about adding appropriate .claude/CLAUDE.md and .cursor/ files and check them in? People report substantial changes in the output of AI when prompted appropriately. Having a more neutral PROMPTING.md helps with allowing user-level control I suppose but then one must trust the users to actually incorporate it into their workflows.

[franciscojavierarceo]

what do you think about adding appropriate .claude/CLAUDE.md and .cursor/ files and check them in?

+1

I think this would be a great way to force some consistency across the bots if everyone's using them anyways.

[rhuss]

Do Claude and Cursor use different formats? Or would it make sense to reuse one file for both locations? (could be done by automation or manually). I'm happy either way, but having the rules like ("Don't touch anything outside the task at hand!") can be added everywhere.

But it's a good idea to have those files that automatically can get picked up by the most prominent AI-assistant tools (not sure what Codex is using or the Gemini CLI).

[rhuss]

Looks like that other projects are also experimenting with adding different LLM-context documents into the tree.

E.g. here is what Kubernetes is currently discussing around adding an AGENTS.md --> kubernetes/kubernetes#133386

Since there are so many different files for different tools, but with the same content, I wonder whether if we could settle on a single file (like PROMPTING.md but the name does not really matter), and then add CLAUDE.md, .cursor/rules/prompting.md and AGENTS.md with the only content only (adapt accordingly):

### Read per-directory guidance and conventions
 
Always look for a `PROMPTING.md` file in each directory, and all of its
parent directories.  Any rules in those files must be followed as it were
rules directly added in `CLAUDE.md`. The closer the `PROMPTING.md` file is to
the code in question, the more priority it should have.

If this sounds like a good idea, I can drive a PR for adding this.

[ashwinb]

After thinking through this, I think we should actually add at least the following files:

• .claude/CLAUDE.md
• .cursor/<whatever cursor needs>
• AGENTS.md -- don't know who uses it

Each agent requires different instructions formatted differently since they use it in various ways. Creating a PROMPTING.md leaves us in a half-way place not accomplishing the final result. All these agents allow users to specify additional rules via their home directories. These are project-specific rules we want to have control over.

Thoughts @ehhuang given you have been experimenting with Claude recently?

[ehhuang]

Agree that we should definitely add these. They will at least offer some consistency with among people's AI usage.

[rhuss]

Actually AGENTS.md is use by OpenAI Codex, but also other tries to establish as a standard, e.g. https://ampcode.com/AGENT.md

There is also the idea to use a single source of truth for the AI-assistant codex like in the document above:

# Cline
mv .clinerules AGENT.md && ln -s AGENT.md .clinerules

# Claude Code
mv CLAUDE.md AGENT.md && ln -s AGENT.md CLAUDE.md

# Cursor
mv .cursorrules AGENT.md && ln -s AGENT.md .cursorrules

# Gemini CLI
mv GEMINI.md AGENT.md && ln -s AGENT.md GEMINI.md

# OpenAI Codex
mv AGENTS.md AGENT.md && ln -s AGENT.md AGENTS.md

# GitHub Copilot (replace [name] with your actual filename)
mv .github/instructions/[name].instructions.md AGENT.md && ln -s ../../AGENT.md .github/instructions/[name].instructions.md

# Replit
mv .replit.md AGENT.md && ln -s AGENT.md .replit.md

# Windsurf
mv .windsurfrules AGENT.md && ln -s AGENT.md .windsurfrules

I don't mind having multiple files, but how can we ensure that all those files are consistent and contain the same content? Also, I don't believe they have different syntaxes; don't they just add the rules from those files directly to the query context? (but I can be wrong, haven't really looked into that)

BTW, as mentioned above, Kubernetes is heading towards a project-wide AGENTS.md (kubernetes/kubernetes#133386)
Anyway, happy with any way you decide to go, as I believe it could make the lives of reviewers a little bit easier.