Show HN: Capshelf – Share agent skills across repos with per-project lockfiles

The Show HN post introduces Capshelf, a tool for sharing Claude Code agent skills across repositories while pinning each project with its own lockfile. The author was running several projects in parallel and had copied the same skills into each repo. Over time those copies drifted: a security-review skill kept improving in one place never updated elsewhere, and it became unclear which copy was current.

The technical pressure is the tension between duplication and global sharing. Full copies per repo avoid cross-project surprise but diverge. Symlinking into a shared directory removed the copies, yet any edit for one repo immediately changed agent behavior in every other repo that pointed at the same files. Capshelf’s stated approach is to share skills across repos while recording per-project lockfiles so each project can depend on a controlled skill set without silent global mutation.

For engineers and builders who live in multi-repo workflows, agent skills are becoming configuration that shapes reviews, refactors, and security checks. When that configuration is ad hoc per clone, skill quality and safety reviews become non-reproducible. When it is a single live shared folder, one local tweak can change behavior on unrelated work. Either failure mode makes it hard to trust what the agent will do on a given checkout.

The market context is the rise of coding agents that load project-local skill packs. Teams already version app code carefully; skill definitions often still sit outside that discipline. Capshelf sits in the same problem space as package managers and lockfiles for libraries: share a common library of behavior, but pin each project so upgrades are explicit rather than ambient.

The practical takeaway is to treat agent skills like any other shared dependency. Prefer a share-plus-lockfile model over either raw multi-repo copies or bare symlinks if you maintain the same Claude Code skills in more than one project. Watch whether Capshelf’s lockfiles make skill upgrades deliberate, auditable, and reverseable per repo without reintroducing silent cross-project drift.