scrxde
16.5K posts





Something we've been trying out the last few days: instead of prompt stuffing AGENTS.md (especially in subfolders), add skills to the repository. So far the results seem better. Its totally possible its entirely behavioral ("just tell it to use the skill" vs assuming it does the right thing), but thats actually fine. Two things that we've done that I think are successful: 1) A bunch of tedium has gone into global skills (github.com/getsentry/skil…) 2) Specific skills in repos that follow patterns ("ui-design", "db-migration", "write-docs", etc.. github.com/getsentry/abac…) Then instead of throwing everything into AGENTS.md, its a lot simpler to just refer it to skills ("when doing X, use Y skill). It requires you to treat the skills like you'd treat any other code investment: maintain and distribute them. This is why I think its also important to do project-scoped configuration for everything (note: see the claude-settings-audit skill). You really need to build these once for the team (or once for the project), and maintain and iterate on them when they have problems. Anyways, so far so good. Nothing ground breaking - just an easy way to share prompt templates - but its smoothing out a lot of silly time sinks. I really want to see some of these behaviors become more cross-agent capable. Cursor's agent will pick up the in-repo skills, but the Claude plugins are specific to Claude right now.

















