A reusable skill that forgets an investigation's scope is incomplete. Knowing how to run a procedure helps little when its next invocation inherits vague permission. I want to reuse the capability without reconstructing from memory the conditions that made using it acceptable.
The case was delegated research I wanted to turn into a skill, to take a capability out of improvisation and be able to call it again. The day before, OpenAI had detailed its Daybreak expansion, which pairs cyber capability with access verification and containment. The same question shows up at a much smaller scale when I delegate an investigation while the main work continues. When a capability concerns security, its instructions need to specify where to look. Knowing a technique does not grant an agent permission to use it on every system where it could help.
I would begin such a skill with the authorized targets and expected result. The environment needs to prevent an attempt from going beyond those targets. Copying a procedure from an earlier assignment requires reviewing that part, because the old authorization may have ended with the work. The procedure stays useful, and the permitted destination may be a different one.
I wrote about autonomy within limits in the July 28 post, "The budget ran out. Where did the agent stop?". Here it means separating reading and investigation from applying a change. Both stages can be authorized at the start. Where they are not, the research ends in a reviewable proposal with the boundary clearly identified, the observations behind each suspicion and the reason for any limitation, so somebody can review it without repeating the whole search. Returning a generic question after every discovery just hands the coordination work back to me. The agent must be able to end an attempt without taking extra actions to manufacture certainty, and that exit belongs in the reusable capability itself.
My criterion for turning research into a skill is that the next execution receives a clear scope and returns understandable evidence, without me remembering on every call which access applies and what counts as finished. In practice, that is a short block at the top of the skill, written before any procedure step: "targets: repo X, read-only; done = proposal with evidence".