A 15% skill-firing rate is not a feature; it is a leak in the user's trust. When a software tester published results showing that Claude Code's six testing skills fired correctly only 15% of the time, and that pushy skill descriptions added a mere five points, the instinct is to blame the model. But the real culprit was the interface. A simple session-start hook, a structured prompt that preceded every interaction, lifted accuracy to 62%. That is not a model fix. It is a design lesson.
This story matters because it reveals something uncomfortable about how we build with AI. The tester found that near-miss requests, queries that were close to a skill's intent but not exact, never fired a skill at all. The system was brittle, waiting for perfect phrasing. This echoes a pattern we have seen before. In When Simpler Tools Match Advanced AI on Intent Classification, we observed that Jev, a tool requiring no training and at most three examples per intent, matched or beat more complex systems on classification. The lesson is consistent: clarity of structure often outperforms raw model capability. A session-start hook is the same principle applied to the conversation itself.
The jump from 15% to 62% is not about making the AI smarter. It is about giving the AI a consistent context from which to operate. For users who feel constrained by traditional spreadsheets or frustrated by AI tools that seem to misunderstand them, the takeaway is direct: how you start matters more than what you ask. This is not a trivial insight. It suggests that many of the accuracy problems we blame on models are actually interface problems. We are asking users to guess the right incantation when what they need is a shared frame of reference established at the outset.
What should you watch for next? The gap between 62% and 100% is still wide. The session-start hook closed the distance, but it did not eliminate misses. That remaining gap is where the next innovation lives, perhaps in adaptive hooks that learn from user behavior, or in hybrid approaches that combine the structure of a session start with the flexibility of near-miss handling. The tester's results are a concrete benchmark, not a finish line. For anyone building or using AI tools, the question is no longer "Is the model good enough?" but "Is the session start good enough?" That is a question worth asking before your next query.