what is infra-skills fr?

Open Infra Skills is a collection of instructions, called agent skills, meant to be loaded by an AI coding assistant so it can help with AI infrastructure work such as training large models, serving them, and optimizing how they run on hardware. Rather than being a library you import into code, it is a set of structured documents an assistant reads to follow a repeatable, tested workflow instead of improvising. The project's stated reason for existing is that AI infrastructure work involves environment specific details, expensive experiments, and performance numbers that are easy to measure incorrectly. Each skill is meant to help an assistant set a reliable baseline before changing anything, pick the right level of diagnostic tool, keep numerical results correct while chasing speed, and clearly separate what was actually measured from what is only a guess. The skills are also meant to work whether the assistant is running on a local machine, over SSH, inside a container, on a job scheduler, or in the cloud. Right now the repository ships one skill, for profiling, debugging, and optimizing AI training on Moore Threads MUSA GPUs, covering utilization metrics, the Moore Perf profiling tool, flash attention, PyTorch's compile feature, fully sharded data parallel training, and the MCCL communication library. A machine readable catalog file lists all available skills so tools can discover them automatically. Each skill lives in its own folder with a main instructions file plus optional agent specific metadata, background reference documents, and helper scripts. The project describes planned but not yet built categories covering distributed training, inference serving, compiler and kernel work, data pipelines, monitoring and benchmarking, and cluster operations, and is explicit that these are a direction rather than a promise of empty placeholder folders. Every skill added to the repository is expected to state real, actionable steps rather than a loose list of tips, include ways to verify the result, and avoid embedding credentials or private infrastructure details. The project is in early development and is released under the Apache 2.0 license.