Packsmith

Build once, ship everywhere for agent skills.

packsmith is a zero-dependency CLI for turning a reusable skill pack into target-specific bundles for coding agents like Claude Code and Codex. It exists because the agent-skills ecosystem is exploding, but the authoring, validation, packaging, and install surfaces are still fragmented.

Why this repo now

As of May 13, 2026:

• GitHub's trending page describes itself as what the community is "most excited about today", and agent tooling dominates the page.
• GitHub's agent-skills topic lists 4,526 public repositories.
• anthropics/skills shows 133k stars and an active issue queue around discoverability, duplicate loads, and install-path mismatches.

That combination is the signal: the category is proven, but the workflows are still rough.

Positioning

Packsmith is not another giant skill library.

It is the packaging layer between:

• teams writing reusable skills/prompts/workflows
• agent runtimes that all expect slightly different layouts
• developers who need validation, portability, and context-budget discipline

MVP

The first version ships five commands:

• init: scaffold a new pack
• validate: verify a pack manifest, skills, and prompts
• inspect: summarize targets, assets, and estimated context weight
• build: compile one source pack into target bundles
• install: copy a built target bundle into a local destination

Current targets:

• claude-code
• codex

Quick start

node src/cli.mjs init my-pack
node src/cli.mjs inspect examples/research-launchpad
node src/cli.mjs validate examples/research-launchpad
node src/cli.mjs build examples/research-launchpad
node src/cli.mjs install examples/research-launchpad/dist/research-launchpad --target codex --dest .local-packs

Built output lands in dist/<pack-name>/.

First-run flow

node src/cli.mjs init my-pack
node src/cli.mjs validate my-pack
node src/cli.mjs build my-pack
node src/cli.mjs install my-pack/dist/my-pack --target claude-code --dest .demo-install

That gives the repo a real story:

1. create a pack
2. validate it
3. build it for target runtimes
4. install the built bundle locally

Pack format

Each pack is just a directory with a packsmith.json manifest:

{
  "name": "research-launchpad",
  "description": "GitHub demand research pack for coding agents.",
  "version": "0.1.0",
  "targets": ["claude-code", "codex"],
  "skills": [
    {
      "id": "github-demand",
      "path": "skills/github-demand",
      "description": "Research GitHub demand signals and convert them into a launch thesis."
    }
  ],
  "prompts": [
    {
      "id": "launch",
      "path": "prompts/launch.md",
      "description": "Turn raw GitHub signals into a concrete MVP."
    }
  ]
}

Repository layout

.github/
  workflows/ci.yml
docs/
  github-demand-2026-05.md
  launch-plan.md
  positioning.md
  specification.md
examples/
  research-launchpad/
src/
  cli.mjs
  lib/
tests/

Engineering posture

• zero runtime dependencies
• Node 20+
• CI runs npm test on pushes and pull requests
• example-first development with a bundled research pack

What makes this star-worthy

• It sits on a real distribution bottleneck instead of shipping one more prompt bundle.
• It is agent-agnostic by design.
• It pushes a simple product story: author once, validate once, build for every agent.
• It gives the ecosystem something missing today: a portable pack compiler with context-aware inspection.
• It now has a complete local loop instead of stopping at compilation.

Next steps

• add install adapters for local user directories
• add lockfiles and remote pack sources
• add target-specific lint rules
• add pack registry metadata and signed releases

Maintainer docs

• Positioning
• Launch plan
• GitHub demand research

License

MIT

Sources

• https://github.com/trending
• https://github.com/topics/agent-skills
• https://github.com/anthropics/skills
• anthropics/skills#675
• anthropics/skills#919
• anthropics/skills#1101
• anthropics/skills#278