Ivan Charapanau

You can run local LLMs with Harbor and llama.cpp, build Python Boost modules for LLM proxy pipelines, orchestrate coding agents in Docker sandboxes, automate bug hunting and fixing, manage SQLite-compatible Turso databases, and develop tinygrad deep learning models.

Developers and ML engineers running self-hosted LLM infrastructure benefit most. The skills target practitioners managing local model serving, containerized agent execution, integration testing, and codebase quality enforcement rather than hosted cloud API consumers.

Use the harbor and run-llms skills to manage containerized LLM services with Bash and Docker commands. The pull-llamacpp-model skill fetches models with automatic ROCm or CPU image selection, and test-boost-module validates outputs through llamacpp.

The discipline skill enforces 15 failure-prevention rules distilled from 120+ real agent sessions. It covers fabrication prevention, constraint tracking, verification, scoping, retry discipline, and communication, and should be loaded before any task to prevent common agent failure modes.

Most skills require Docker for containerized services and sandboxes, Bash for Harbor management, and Python for Boost module development. GPU-accelerated model serving uses ROCm images, while frontend work with preact-buildless-frontend needs only a browser supporting ES Modules.