Hi HN,

I was getting incredibly frustrated with the current state of AI agent skill managers (like skills.sh). They rely on heavy Node.js runtimes just to manage a few skills, symlinks break constantly across environments, and there's no real determinism.

I wanted a tool that treats agent skills like Terraform treats infrastructure: Config-as-Code, purely deterministic, and zero-dependency.

So I built eden-skills — a single ~10MB Rust binary based on Tokio that uses a skills.toml to lock every skill by commit SHA. It features a doctor/repair command that actually self-heals broken symlinks and drifted states, and it can inject natively into Docker containers (--target docker:<container>).

But the real experiment here is how it was built.

Instead of writing it all myself, I acted purely as the orchestrator for a 3-model AI team, using a strict Role-Based Access Control (AI RBAC) approach to eliminate context loss and hallucination:

- The Scout (Gemini 3.1 Pro): Handled market research and roadmap drafting.

- The Architect (Claude Opus 4.6): Wrote 62 behavior specs (11k+ lines of markdown). I used strict prompt constraints to explicitly forbid Claude from writing any implementation code.

- The Builder (GPT 5.3/5.4): Implemented the Rust code directly from the frozen specs. It was forbidden from altering the design or modifying the specs.

By physically isolating their permissions through distinct kick files and a batch-handoff protocol, this pipeline generated ~38,500 lines of Rust and 456 automated tests—all passing.

For me, the most valuable part of this project isn't just the CLI tool itself, but the reproducibility of this workflow. I've open-sourced the entire prompt/ directory containing all the system prompts, role constraints, and handoff protocols I used to manage the models.

If you are building deterministic systems, tired of JS-heavy CLI tools, or just curious about how to orchestrate multi-model development without things catching on fire, I'd love for you to tear the architecture apart or check out the workflow.

Repo: https://github.com/AI-Eden/eden-skills

Agentic Workflow Docs: https://github.com/AI-Eden/eden-skills/blob/main/docs/agenti...

Prompts: https://github.com/AI-Eden/eden-skills/tree/main/prompt

Happy to answer any questions about the Rust implementation (two-phase async reactor) or the AI orchestration process!