Filter your skill/tool list by what the account state can actually execute, not by semantic relevance: 90.5% context reduction in production across 756.6K messages
Wix's customer-care assistant Helpmate ships a three-stage selection pipeline where a deterministic executability gate sits between semantic matching and the LLM's choice, removing any skill whose internal hard-stop conditions currently hold. In a post-launch analysis of 756,600 user messages across 267,600 conversations, the gate removed 1,039,462 of 1,749,270 skill-message pairs (59.4%), saving 228.8 million skill-description tokens; combined with semantic matching that is a 90.5% cut in skill-description context versus exposing all ten skills every time. The counterfactual is the part that matters for builders: replaying a risk-enriched cohort of 1,000 conversations with everything exposed, the model picked a production-blocked skill in 7.8% of them — so the gate is not just a token optimization, it removes choices the model would otherwise make wrongly.
↳ Follow the thread