arle

Use this skill when ckl asks to ground an ARLE serving/runtime, model-path, benchmark, capacity, Qwen3.5/DeepSeek, scheduler, paged_kv, MLX, autograd, or OPD decision in upstream SGLang/vLLM/TensorRT-LLM evidence before changing local code. It distills BBuf AI-Infra Auto Driven SKILLS into an ARLE-specific source-survey workflow without symlinking or vendoring those repositories.

Use this skill when ckl asks to optimize an ARLE kernel, operator, attention path, GEMM, decode/prefill path, quantization op, scheduler hot path, TTFT/ITL/tok-s metric, memory footprint, or any "optimize this operator" / "tune the kernel" / "make this faster" request. It enforces formula-predict -> measured binding constraint -> matched single-variable A/B -> interaction A/B when needed -> explicit tradeoff -> license-or-kill, and keeps the industry catalog scoped to ARLE CUDA/TileLang/MLX/runtime work.

Use this skill when ckl asks to inspect, queue work for, interrupt, spawn, replace, or otherwise drive another coding-agent CLI running inside tmux. Covers ARLE's known-safe tmux path for Codex/Claude Code delegation, including session discovery, capture-pane status checks, Enter semantics, long-brief buffer paste, queue-vs-immediate behavior, and don't-send-to-yourself safety.

Use this skill BEFORE writing code for any task you're tempted to call "hard / tough / complex", any "why is X slow / where's the bottleneck", any "should I optimize Y", any concurrency/perf/kernel/scheduler investigation, or any time you catch yourself hand-waving a root cause. It is the pre-implementation understanding GATE — decompose to the atomic (line/kernel/buffer) level, get MEASURED evidence, let measurement correct your hypotheses, until the problem is SIMPLE and quantified. Only then write code. If it still feels hard, you haven't decomposed enough.