The highest-scoring skill in our productivity category doesn't produce anything. It asks a question. After evaluating 600+ agent skills, we think that's not an anomaly — it's the most load-bearing thing a skill can do.
We run Skill123, a directory where every agent skill gets scored on six dimensions with written rationale. Sort by score and past the usual suspects — Office tooling, API references — a pattern appears that took us sixty evaluations to name: the top of the list includes skills whose entire mechanism is making the agent stop and ask, at the exact moment where guessing would be expensive.
Three mechanisms, three skills:
1. The pre-answer check
academy-guide (9.6, our productivity #1): "Stop and check this skill before finishing any reply." It encodes a discipline — verify against a standard before committing to an answer. It's the closest thing we've seen to a skill implementing self-review.
2. The extract-want
interview-me (9.4): "Extracts what the user actually wants instead of what they tell you." Users describe solutions, not problems — "add a cache" when the problem is a slow query. The skill's entire content is a set of questions that separate the stated request from the underlying goal, each with a decision attached to the answer.
3. The post-answer nudge
discernment-nudge (9.5): intervenes after the answer, at the moment you're about to act on it. Most skills are silent there. That's exactly when overconfidence costs the most.
Same mechanism in marketing: linkedin-interviewer (9.4) interviews the user for post raw material — the "I don't know what to post" user gets asked questions until they do.
Why this scores so well
On the surface these look trivial — a question is not a capability. But the real bottleneck of agent work isn't capability, it's specification. An agent with a mediocre skill and a correct spec beats an agent with a brilliant skill and a vague one. Interview skills manufacture correct specs.
There's also a scoring reason: in our data, the dimension that decides whether a skill ever helps is trigger quality — knowing when to fire. Interview skills take the opposite risk and manage it deliberately: they fire on every substantive interaction, and their entire content is about what to do in that moment.
The recipe, if you want to write one
- Pick the moment, not the domain. Before an answer, after an answer, before a deploy — the timing is the skill.
- Write the questions, not the answers. Each question should have a decision attached; that's what makes it sharp rather than chatty.
- Bound the interruption. Say when not to ask, or the user will uninstall you.
- One nudge, not a form. Ten questions is a survey, and a survey is a different (worse) product.
The test: after the skill runs, does the agent's next action have a better chance of being what the user wanted? If yes, the skill worked — no matter how little it "does."
The full pattern breakdown — including the LinkedIn pair and how these skills bound their triggers — is at skill123.me/blog/one-question-interview-skills. Every skill mentioned is installable from its profile page.
Top comments (0)