Skill++: Helping AI agents find the right skills as their libraries grow

Adding capabilities to an AI agent can make its existing capabilities harder to find. When several skills have similar names or descriptions, a newly installed skill may divert requests from an older, more appropriate one. Skill++ tackles this access problem by refining how each incoming skill is presented, rather than retraining the model that chooses it.

When more skills create confusion

An AI agent’s skill library contains reusable packages of instructions, code and reference material. Each package has a name, a short description and a body. To avoid loading every package into its limited working context, an agent can first choose skills using only their names and descriptions, then open the selected bodies. This arrangement is called progressive disclosure.

The difficulty is that descriptions do not operate in isolation. A label that clearly identifies one skill in a small collection may become ambiguous when similar packages arrive. The earlier skill has not changed, and neither has the model choosing it, but the agent can nevertheless stop selecting it correctly.

In the Skill++ preprint, Hanrong Zhang and colleagues call this problem “lifelong skill routing.” They distinguish two forms of interference: substitution, where a similar newcomer replaces a required skill, and crowding, where other skills fill the few available selection slots.

Editing the entry, not retraining the router

Skill++ changes only the name and description of an incoming skill. It leaves earlier entries and all skill bodies unchanged, and does not train the router or require labeled routing examples.

The method has three components. First, it finds the earlier skill with the most similar original name and description. A language model reads passages from both bodies and writes a new description emphasizing the incoming skill’s distinctive capabilities. The prompt requires those capability claims to be supported by its body.

Second, Skill++ adds short exclusions describing nearby work the incoming skill does not cover. To find relevant neighbors, it combines similarities between metadata and synthetic user queries generated from skill bodies. These generated queries support the neighbor search; they are not labeled evaluation examples.

Third, when a name overlaps with an earlier name, the method can add a qualifier identifying the topic or workflow. Original skill records remain available for future comparisons, so later refinements do not rely on a chain of previously rewritten descriptions.

What the main experiment measured

The researchers adapted three existing benchmarks into growing-library tests. Their main study used 198 SRA-Bench queries derived from BigCodeBench, each requiring three to five skills. Across ten sessions, the library expanded from 27 to 731 skills.

Incoming candidates were deliberately selected to resemble earlier required skills, appear confusable from metadata, and fail a model-based check of whether they could replace the required skill. This makes the main experiment a targeted test of interference, rather than a random sample of installations.

The main router, Claude Haiku 4.5, shortlisted ten skills and then reordered them. The researchers scored the first five. Their primary measure, full coverage at five, counts a query as successful only when every annotated required skill appears in that selection. It measures access to skills, not whether the agent ultimately completes the task.

At the final session, the authors report full coverage of 76.3% with Skill++, compared with 36.4% using unchanged metadata. The strongest comparison method reached 67.8%. Removing any of Skill++’s three components reduced coverage, with description differentiation contributing the most.

Wider results and their boundaries

Across five routers, the reported final-session improvement over unchanged metadata averaged 23.1 percentage points. Skill++ also improved coverage on SkillRet and SkillRouter as their libraries approached 2,000 skills. However, final coverage there remained 30.0% and 38.7%, respectively: improvement did not eliminate routing failures.

The authors also tested native skill files inside Codex CLI, increasing its skill-list budget so the full library fit. Selection improved with both evaluated models. A separate task-completion check used 150 queries and the smaller, 87-skill library. Hidden-test pass rates increased by 4.0 and 0.7 percentage points, showing why selection gains should not be equated with equally large execution gains.

The study is limited to textual skills, queries and router inputs; the authors leave images, audio and other modalities to future work. Its central claim is narrower than creating new agent abilities: clearer, more differentiated metadata can help preserve access to capabilities already present as a library expands.