Max Throughput — Development skill for Claude Code

max-throughput

An [Agent Skills](https://agentskills.io) compatible skill that makes coding agents (Cursor, Codex, Claude Code, ...) stop running your expensive machine single-threaded.

**The problem it solves:** agents happily execute `python train.py` on an 8-GPU box and `pytest` on a 128-core workstation, one thing after another. The default objective becomes *minimum wall-clock completion time*: probe the hardware, build a work DAG, and dispatch everything concurrently - agent-level parallelism, multi-GPU DDP, per-GPU experiment packing, `pytest -n auto`, `make -j$(nproc)` - then verify utilization and escalate parallelism if resources sit idle.

Install

With the [skills CLI](https://skills.sh) (works for Cursor, Codex, Claude Code and more):

# pick your agent with -a, repeat to install for several at once
npx skills add Promethe-us/max-throughput -g -a cursor
npx skills add Promethe-us/max-throughput -g -a codex

Manual install: copy this folder into your agent's skill directory, e.g.

~/.cursor/skills/max-throughput/
~/.codex/skills/max-throughput/     (or reference via AGENTS.md)
~/.claude/skills/max-throughput/

What the agent will do differently

Before	After
python train.py on 8 GPUs	torchrun --standalone --nproc_per_node=8 train.py
experiments queued one by one	experiments packed one per GPU via CUDA_VISIBLE_DEVICES
pytest	pytest -n auto
make	make -j$(nproc)
serial preprocessing loop	sharded inputs via multiprocessing / GNU parallel
guesses a knob when slow	profiles first (quick_triage.py, py-spy, torch.profiler), then tunes the measured bottleneck: batch size, num_workers, prefetch, pre-encode, AMP/compile
never checks anything	runs scripts/inspect_resources.py, then nvidia-smi / top to confirm the machine is actually saturated

Structure

max-throughput/
├── SKILL.md                        # the skill itself (name + description + instru