Tools: Code Is All You Need

Using LLMs with CLI Snippets vs MCP Tools

• Several commenters report strong success with simple “playbooks” (e.g. CLAUDE.md) full of shell commands and examples. The LLM learns patterns from these and reliably adapts them to new, similar tasks.
• Others note you can often turn such command collections into very thin tools (e.g. MCP servers or scripts) but question whether that adds meaningful value over terminal access plus good instructions.
• Some argue MCP shines mostly for poorly documented, proprietary, or internal systems where you can hide auth/edge cases behind a stable tool interface.

Context, Composition, and Scaling Limits of MCP

• A recurring complaint: every MCP tool definition consumes context. With many tools, “context rot” degrades performance; some report practical limits of ~15 tools.
• MCP is seen as less composable than shell pipelines: each tool call is separate, with intermediate data routed via prompts instead of native pipes.
• Others counter that tool schemas plus constrained decoding reduce errors versus free-form command generation, though skeptics say the gain is modest.

Reliability, Safety, and Sandboxing

• Many participants are uncomfortable letting LLMs directly touch production systems; they prefer tools as a permission/constraints layer, or have the LLM propose commands for human review.
• Sandbox patterns (VMs, Docker, read-only mounts, language REPLs like Julia/Clojure) are popular; they noticeably cut token usage and make LLMs more likely to reuse existing code.
• Some note that autonomous “agentic” setups still underperform guided, human-in-the-loop workflows.

Economics, Hype, and Appropriate Use

• Multiple comments compare LLM hype to 3D printing, VR, drones, NFTs, and the Metaverse: useful but far narrower than maximalist predictions, with unresolved business models and heavy infra cost.
• Others push back, pointing to widespread everyday use (especially ChatGPT) and seeing LLMs as a real paradigm shift, especially for translation, research, and coding assistance.
• There’s concern that subscription prices and rate limits will rise as subsidies fade; some expect open or local models to catch up enough for many coding tasks.

Shell vs Higher-Level Languages

• Strong divide over bash/Unix CLI: some see it as the perfect universal substrate for LLM-driven automation; others find the ecosystem archaic, error-prone, and unusable on Windows, preferring Python or other languages as the “script target” for code generation.