Post by Elastic

537,945 followers

80% of agent skills yield zero improvement over no skill at all. Some actively hurt performance. That's not a niche edge case — that's the default outcome if you write skills without understanding what actually makes them work. The failure pattern is consistent: skills that trigger at the wrong time, skills that are too vague to differentiate tools, skills that add noise instead of context. The agent doesn't fail because the underlying model is bad. It fails because the skill told it the wrong thing at the wrong moment. We walked through the research and built a real example using our ES|QL skill, including what happens when you let an AI model write the skill for you. Worth 15 minutes if you're building agents that need to actually work.

Post content

Video Content