Agentic Distribution

The problem of keeping private agentics — prompts, agents, and skills — synced and un-duplicated across many codebases, devices, team members, and agent fleets. Below a couple of repos it is a non-problem; at 10+ codebases with agents on multiple devices it becomes acute: you copy a skill, it drifts, and there is no shared access. The solution is a meta skill plus a single reference manifest — a YAML "library" file that behaves like a package.json / pyproject for your agentics. Crucially the manifest stores pointers (private GitHub repos or local file paths) to where each artifact actually lives, never copies. A small, package-manager-shaped command surface operates over it: add (catalog a new artifact as a reference), use (install an artifact by reference into a target namespace), push (write a local edit back to the source repo), list / search, and sync (pull the latest source, not just the catalog). Because there is one source of truth and everything is a reference, any device, agent, or teammate that clones the manifest + skill is bootstrapped with the same, latest agentics. The public reference library (just the YAML + skill) is kept distinct from the private artifact repos it points at, so the pointer file can be shared while the valuable, specialized agentics stay private. The distribution tool itself is a Pure Agent Application — a SKILL.md + library.yaml, no executable code.

Claims

Distribute agentics by reference, not by copy — a single manifest points at where each skill/agent/prompt actually lives, giving one source of truth instead of duplicated, drifting copies. principle — durable: copying is the mechanism that produces drift; a pointer to a canonical location is what keeps N surfaces in sync by construction.
Bootstrap any device, agent, or teammate from two small artifacts — the reference manifest and the meta skill — rather than shipping the agentics themselves. best practice — context: an operator/team running specialized agentics across 10+ codebases, multiple devices, and agent fleets; below that scale (1–2 repos) it is not worth it — "just use plugins / install from wherever."
Split the shareable reference library from the private artifact repos it points at, so the pointer file can be public while the specialized agentics stay private. best practice — context: you build (and get paid for) specialized agentics; top-notch skills/agents/prompts stay in private repos and only the manifest is distributed.
Give distribution a package-manager command surface — add (catalog), use (install by reference into a namespace), push (write edits back to source), list/search, sync (pull latest source). best practice — context: managing agentics at scale; push is what lets any device edit-in-place and propagate the change to the source, and sync pulls the actual referenced code bases, not merely the catalog file.
Skip versioning — for internal, fast-moving agentics you almost always want the latest, so sync-to-latest beats pinned versions. best practice — context: internal/team agentics with a trusted single source; a published, externally-consumed package would instead want real versioning, so this is contingent on "one owner, always latest."
Install by reference into a chosen namespace — local project vs. global (~/.claude) vs. arbitrary named routes — pulling artifacts from the source repo at install time. observation — the same referenced artifact can land locally in one device and globally in another; the manifest carries configurable default/global/special destination routes.
The reference-not-copy principle is a claim about your agentics; third-party consumption is the boundary case it concedes, and the concession has an unclosed edge. observation — a second source documents the mass-adoption path in the wild: git clone <public-repo> ~/.claude/skills/<name>, a copy by construction. That is exactly the "1–2 repos, just install from wherever" case above, so it does not falsify the principle — but the drift it predicts is real and unaddressed. One of the cloned repos was at version 3.8.1 with a commit five days old; a cloned copy is stale on landing and nothing re-pulls it. Consumer-side updates are the seam between this concept and Public Skill Adoption, and neither side currently owns them.

Reusable Workflow Library — the sibling concept (catalog reusable agent workflows); agentic-distribution extends it with the reference-manifest + cross-device sync/push mechanism and broadens the payload from loops to skills+agents+prompts.
Layered Agentic Architecture — the primitives being distributed (skill = capability, agent = scale, prompt = orchestration, task runner on top) are exactly this stack; "agentics" is the umbrella term for them.
Meta-Prompt — meta-agentics (skills that build skills/prompts/agents) are the artifacts most worth distributing identically across every device, teammate, and agent.
Pure Agent Application — the distribution tool is one: it runs entirely in SKILL.md + library.yaml, no code.
Agent Supervision — references-not-copies over open plain files is the transparency / "know exactly what your agents run" angle applied to the agentics themselves.
Spec-Driven Development — the same "share the definition, not the rebuilt copy" instinct, one altitude down (share the reference, not the duplicated artifact).
Public Skill Adoption — the consumer mirror of this concept: acquiring other people's agentics by clone-and-copy. It is the boundary this page's best-practice concedes; the shared unsolved problem is what happens to an installed copy when upstream moves.
Distillate: The Library Meta-Skill: How I Distribute Private Skills, Agents and Prompts.
Distillate: 160,000+ Cloned These 3 FREE AI Employees: Here's How (GitHub Claude Skills)

Agentic Distribution

Claims

Related

Linked from