One canonical deck of skills, agents, and rules, compiled for every AI harness. One source of truth, four native dialects.
Most teams scatter their agent instructions across four dotted directories and let them rot independently. A deck keeps one canonical source, written once in portable markdown, and compiles it for Claude Code, Codex, Gemini, and OpenCode. The deploy is three commands long.
$ rune add core --source ~/decks/runedeck staged rune selection 'core' → .rune $ rune install ● 137 deployed $ rune doctor --target . ./.claude/skills/rune (claude) ok 31 · modified 0 · missing 0 · orphan 0
Canonical markdown with Agent Skills frontmatter. Provider variants merge at assembly: argument hints for Claude, TOML agents for Codex, remapped tools for Gemini. Your instructions become a build artifact.
Every deployed file carries SLSA provenance back to its source line. doctor never overwrites your edits, drift shows exactly what moved, and validation names what it could not check.
Third-party skills enter through block-by-block review with injection linting. Every block gets a verdict: keep, adapt, or cut. Nothing lands unverdicted.