## Does agent_skills support automatic skill selection based on task context?

[marcorobiati]

I have multiple mutually exclusive skills configured for gsd-executor — e.g. a frontend-react-coder skill and a backend-spring-boot-coder skill. Both are assigned in agent_skills and would be injected into every executor prompt regardless of what the task actually involves.

My assumption is that since GSD injects skills as @path/to/SKILL.md references, the subagent reads the full file including the frontmatter description — and should therefore be able to apply the same relevance judgment that Claude Code natively does when selecting skills (i.e. pick the most relevant skill based on task context, and ignore the rest).

Is this already the intended and expected behavior? If so, is it documented anywhere?

If not — i.e. if subagents are expected to apply all injected skills unconditionally — then I'd argue this is a gap worth addressing, since the frontmatter description mechanism already exists precisely to express skill applicability.

Happy to run some tests and report back, or open a formal feature request if this isn't already handled.

[bragabriel]

No, that's not the current behavior. The frontmatter description works as a filter before a skill gets loaded, not after.

Native skill selection happens in two stages: only the skill listing (name + description) is preloaded, and the full SKILL.md body is loaded once the model decides it's relevant and invokes it.

From the skills docs:

"skill descriptions are loaded into context so Claude knows what's available, but full skill content only loads when invoked".

The relevance judgment you're describing happens at that listing stage.

@path references skip all of this. According to the memory docs, imports are "expanded and loaded into context at launch", with no filtering. Same thing for the skills frontmatter field on subagents: "The full skill content is injected, not only the description."

So with both skills in agent_skills, both full bodies end up in every executor's context. The model can still lean towards the relevant one, but that's just model discretion. There's no documented mechanism that uses the description of an already injected skill as an applicability gate, and you pay the context cost of both skills either way.

Some options to get the behavior you want:

• Pick the skill at dispatch time, so the orchestrator decides which one to inject based on the task before spawning the executor.
• Skip preloading. If the executor has the Skill tool available, subagents can still "discover and invoke project, user, and plugin skills" through the normal listing mechanism, which does filter by description.

Since the injection strategy is GSD's design choice and not something Claude Code imposes, I'd say this is a valid feature request on GSD's side. The first option is the more deterministic fix.