01 / The challenge
Where the work began.
Useful AI context can become trapped in a single editor or model, duplicated across repositories, and separated from the evidence that originally supported it. Teams need a maintainable source of truth that can adapt as their tools change.
02 / The approach
AI Hub, end to end.
I built a portable Markdown/YAML content model with adapters for Codex, Claude Code, GitHub Copilot, Cursor, and Junie-oriented workflows. A desktop application and CLI handle discovery, validation, installation, synchronization, sharing, and source management.
03 / Engineering decisions
The choices that mattered.
Separate knowledge from adapters
Personal, base, and project context resolve independently of any one AI development environment.
Keep the evidence connected
Lineage, provenance, source snapshots, and review locking preserve the connection between a reusable capability and its supporting material.
Treat local software as a security boundary
Sandboxing, IPC validation, path-traversal defenses, atomic imports, update integrity checks, and rollback address real failure paths.
04 / The result
What shipped.
- Published 32 desktop releases between July 2 and August 11, 2026.
- Reached recorded validation milestones of 191 core tests, 30 desktop tests, and seven end-to-end journeys.
- Delivered portable project and capability sharing with dependency validation, collision handling, and rollback.
- Built macOS delivery and a Windows packaging foundation; signing and Windows device validation remained separate release gates.
An internal Slalom project. Release and test counts are documented milestones, not claims about adoption or measured productivity gains.
05 / What stayed with me
The tools will change. The team's knowledge, its provenance, and the way people govern it should survive that change.