Andrej Karpathy's Sequoia talk, I agree with most but not this (opens in new tab)

2. install .md skills instead of install .sh scripts.

Thanks, but not this.

My nerves hurt when I see redundant, slow, repeated process.

Use the LLM to compile, not execute.

The right solution is a create-installation-script skill: it understands env differences, generates the install script, validates it against the test framework, and produces something that just works.

Do not make the LLM run the same installation steps billions of times.

Similarly, we now rely on agents to handle more and more workflows, it's better to create a harness to let agents 'compile' repeated work into deterministic code, also make sure agents handle the failures/exceptions to self-evolves the logic.

If curl | bash was bad, this can be so much worse. But it's probably days now, before some big project will introduce this.

And no one will have time to review "plan mode" for this script.