Select blind targets
Choose a small, varied batch whose active skills plausibly serve individuals or small teams. Prefer repositories with a few focused skills over prominent marketplaces, large catalogs, and repositories already selected because their results are known. Explain fit using the observed number and purpose of active skills. Treat stars as a popularity signal and avoid claims about repository size unless measured.
Inspect only enough repository metadata and skill inventory to establish fit.
Before setup or inference, write a selection artifact containing the repository
URL, exact target revision, active SKILL.md paths, selection reason, and any
access or freshness limitation. Do not include prior SkillRoll outcomes or a
prediction about which target will pass.
For revisions taken from a local cache without a network check, state that
remote freshness is unknown because it could not be checked.
When a rejected candidate's recursive count is inflated by fixtures or vendored
examples, record that distinction explicitly so the handoff does not repeat the
misleading raw count.
Read selection context for the evidence boundary and artifact fields.