Update README.md

README.md

@@ -26,9 +26,12 @@ tags:
 
 ## 📌 Overview
 
-**WildClawBench** is an agent benchmark designed to test real-world utility: can an AI agent perform complex work end-to-end without hand-holding? Agents are deployed into a live [OpenClaw](https://github.com/openclaw/openclaw) environment—a real-world personal AI assistant—to tackle **60 original tasks**.
 
-These tasks are designed to be significantly more difficult than existing benchmarks; currently, every frontier model scores below **0.6**, ensuring the evaluation remains meaningful for next-generation agents.
+**WildClawBench** is an agent benchmark that tests what actually matters: can an AI agent do real work, end-to-end, without hand-holding?
+
+We drop agents into a live [OpenClaw](https://github.com/openclaw/openclaw) environment — the same open-source personal AI assistant that real users rely on daily — and throw **60 original tasks** at them: clipping goal highlights from a football match, negotiating meeting times over multi-round emails, hunting down contradictions in search results, writing inference scripts for undocumented codebases, catching privacy leaks before they happen. Useful things. Hard things.
+
+Hard enough that **every frontier model today scores below 0.6**. That makes scores mean something.
 
 ## 📂 Repository Contents