Simon Farshid
313eb74c44
Add one evidence-backed line to AGENTS.md ("When you change code, delete any comment that only records its history.") and the evals/ A/B harness that justified it: an undirected agent runs in a sandbox with/without a candidate sentence, scored by an LLM judge. On the reproduced case, that line takes Opus 4.8 from 0% to ~94%, while vague "write good comments" phrasings stay at baseline.
11 lines
261 B
JSON
11 lines
261 B
JSON
{
|
|
"name": "@assistant-ui/x-prompt-evals",
|
|
"version": "0.0.0",
|
|
"private": true,
|
|
"type": "module",
|
|
"description": "A/B harness for deciding which guidance sentences earn a place in AGENTS.md / skills",
|
|
"scripts": {
|
|
"eval": "node src/cli.ts"
|
|
}
|
|
}
|