If an AI writes down a 'skill' from code, how do you check it's concrete enough that someone could actually reuse it?

This explores how Code2Skill's 'source-blind reconstruction' check works as a test of whether an extracted skill is concrete enough to reuse. The corpus is thin on the mechanics. Can source code replace experience as skill raw material? says the system pulls over a million auditable skills out of GitHub repositories and verifies them this way. It reports a lift from 42.90 to 47.90 macro-average across eight benchmarks, beating skills built from agent trajectories or documents. It doesn't spell out the procedure, scoring, or who does the reconstructing.

The name suggests the logic, though this is my reading, not the note's. 'Source-blind' most plausibly means that someone, probably a model, tries to rebuild what the code does from the skill alone, without seeing the original. A vague skill like 'validate inputs carefully' gives a blind reconstructor almost nothing to work with, so it fails. A skill that names the actual steps, interfaces, and edge cases lets the rebuild succeed. That is a test of reusability, because a skill someone can rebuild from is one that carries enough detail for a stranger to act on. Code is also unusually good raw material for this, because the original still exists as ground truth. That fits the note's point that these skills keep 'executable evidence'.

The corpus suggests why specificity needs an explicit check: skills drift toward vagueness on their own. Can a separate trained curator improve skill libraries better than frozen agents? shows that skill repositories grow generic, verbose additions until a separately trained curator pushes them toward actionable execution logic and cross-task strategies. Reconstruction attacks the same problem earlier, at extraction time. [[task-decomposition-granularity-is-the-primary-bottleneck-in-compositional-skill]} adds a second angle. Standard LLM decomposition reaches only 34% step-level recall, and fixing the step count recovers about 75% of the gains in iterative methods. A skill can only be reused if it sits at a granularity the task can actually match.

Other notes verify skills differently, by using them rather than rebuilding from them. Can skill documents be optimized like neural network weights? gates every edit to a skill document on held-out validation and accepts only changes that improve performance. Can agents learn new skills without forgetting old ones? refines executable skills with environmental feedback. Reconstruction asks whether a skill says enough. These methods ask whether it works. The corpus doesn't compare them, so whether reconstruction predicts downstream usefulness is an open question here. One caution comes from Can attackers evade skill scanners by refining individual skills?: scanners that judge skills one at a time miss what happens when skills are chained. A per-skill specificity check likely has the same blind spot for how skills combine.