Daniel San(@dani_avila7):Excellent paper on how to actually improve a harness Think of a junior dev joining a team, their brain doesn't change, their notebook does... two months in they just make fewer dumb mistakes Same here, the model stays identical, what changes is a folder of notes it loads before each task. Same idea as SKILL.md Then they changed who decides the task failed If the model grades its own work, the notes leave it worse than having no notes at all. If a unit test grades it, it goes up So the reflection was never the valuable part, the verifier is. And that's the loop half of us are shipping right now Link:

Daniel San

@dani_avila7

Head of AI at | Building Claude Code SubAgents, Skills & Hooks | OSS project | Powered by TS, Python & Vanilla Lattes ☕️

가입 January 2015

3.4K 팔로잉 중 33.6K 팬

Daniel San@dani_avila7

2026.08.24 23:15

Excellent paper on how to actually improve a harness Think of a junior dev joining a team, their brain doesn't change, their notebook does... two months in they just make fewer dumb mistakes Same here, the model stays identical, what changes is a folder of notes it loads before each task. Same idea as SKILL.md Then they changed who decides the task failed If the model grades its own work, the notes leave it worse than having no notes at all. If a unit test grades it, it goes up So the reflection was never the valuable part, the verifier is. And that's the loop half of us are shipping right now Link: