What jumped out at me is how small the “pushy” win actually is compared with the folklore around it. People love to talk about prompt wording like it’s the lever that fixes everything. Here it barely moved the needle: five percentage points on the pooled indirect prompts, and even that came from only three of the six skills. That feels much more like “sometimes worth trying” than “this is the trick.”
The more interesting result is the ugly one: near-miss prompts never fired a target skill in any of the setups. Zero. That makes me trust the author’s setup more than a lot of these skill-trigger anecdotes, because near-miss controls are exactly where a sloppy eval tells on itself. If a description with matching keywords still doesn’t trigger on obvious false positives, then Claude Code’s skill selection is probably doing something narrower than simple keyword matching, which is good, but also a bit annoying if you’re trying to design reliable skills.
I also think the session-start hook result is the real story. Jumping from 15% to 62% by telling Claude to check its skills first is not a subtle effect. That’s a much stronger intervention than wording tweaks, and honestly that tracks with how these agents behave in practice: if the system doesn’t explicitly prime the right behavior, the model may just answer directly and never look around. The “pushy description” advice starts to look like a weak local fix, while the hook is closer to changing the operating procedure.
Still, I wouldn’t overgeneralize from this. The test bed was six target skills in a 26-skill workspace, and the author is upfront that the pushy sentence also repeats keywords and changes length, so tone and verbosity are tangled together. That matters. If a longer description works better, is that because it’s pushier, or just because it’s easier for the model to notice? The article can’t separate that, and I think that limitation is real.
What I’d do next, if I were using Claude Code seriously, is treat hooks as the practical tool and descriptions as cleanup. Make the skill descriptions explicit, sure, but don’t expect copyediting them to solve trigger reliability. The article basically convinces me that description tuning is marginal, while a session-start reminder is operationally meaningful.
Reference: A Session-Start Hook Raised Claude Code Skill Firing From 15% to 62%
