A few months ago I realized I dont read my agents thinking anymore.
And models are so keen to please you. Something breaks, they will do whatever it takes to tell you is done. This is not the right approach sometimes.
So I added a skill to my workflow: paper cuts. At the end of every turn, the agent lists what broke outside the task, and I decide what is worth fixing upstream.
npx skills add onmax/skills --skill paper-cuts
And models are so keen to please you. Something breaks, they will do whatever it takes to tell you is done. This is not the right approach sometimes.
So I added a skill to my workflow: paper cuts. At the end of every turn, the agent lists what broke outside the task, and I decide what is worth fixing upstream.
npx skills add onmax/skills --skill paper-cuts