← 输入框|gist.github.com|刷新
Skip to content

Instantly share code, notes, and snippets.

@romansod
romansod / fail-safe-bencmark.md
Last active October 21, 2026 16:31
Exposition

Does a narrower scope converge?

Read the full write-up: https://fail-safe-benchmark.netlify.app/

A small benchmark of a "fail-safe" system prompt for AI coding agents: decide up front what the program supports, and refuse everything else by name, with nothing changed, instead of guessing.

The setup