LLM cmd, an LLM plugin to prompt and edit a shell command (simonwillison.net)

Developer released llm-cmd, an alpha plugin for the llm CLI that asks a large model to generate a shell command from natural-language intent, pre-populates your terminal with that command so you can edit it, then runs it when you hit Enter (or cancels with Ctrl-C). It’s aimed at speeding up common shell tasks (e.g., “undo last git commit” → git reset --soft HEAD~1) and reducing cognitive friction for devs who prefer natural-language prompts. The author stresses it’s dangerous and intended only for users fluent with the terminal — generated commands are executed locally and can delete or modify data. Technically, llm-cmd uses the llm framework (installable via brew or pipx; you must configure API keys) and supports multiple models (gpt-3.5-turbo, GPT-4, Claude 3 variants). The system prompt forces the model to return only a raw command string — no markdown or commentary — because the plugin passes the result directly to subprocess.check_output(). The UX trick is readline.set_startup_hook(lambda: readline.insert_text(command)) to prefill the input buffer so users can edit before execution. Flags include -m to pick a model and -s/--system to change the system prompt. The release demonstrates how LLMs can be tightly integrated into developer tooling but also highlights security and correctness risks that merit caution and iteration.