skill-builder

Publish and install AI skill prompts from the command line — or author them in the web editor with help from an AI drafting assistant.

npx @dmzagent/skill-builder install @author/skill-id

Getting started: installing skills

Browse the registry

Results print as a dense table right in your terminal:

npx @dmzagent/skill-builder list

Filter by category, author, type, tag, and sort order:

npx @dmzagent/skill-builder list --category "Developer Tools"
npx @dmzagent/skill-builder list --type meta          # only meta skills
npx @dmzagent/skill-builder list --author skillauthor
npx @dmzagent/skill-builder list --sort downloads
npx @dmzagent/skill-builder list --json               # machine-readable
npx @dmzagent/skill-builder list --web                # open the browser registry instead

Search and inspect

npx @dmzagent/skill-builder search "data pipeline"    # ranked by relevance
npx @dmzagent/skill-builder info @skillauthor/dialogue-flow
npx @dmzagent/skill-builder suggest dial              # autocomplete: skills, tags, authors

Wire up shell completion (skills, tags, and authors complete as you type):

npx @dmzagent/skill-builder completion >> ~/.bashrc

Install a skill

npx @dmzagent/skill-builder install dialogue-flow
npx @dmzagent/skill-builder install @skillauthor/dialogue-flow

Downloads dialogue-flow.md and dialogue-flow.json to the current directory. Use --output ./my-skills to change where files land.

Meta skills install their dependencies automatically. Installing a meta skill resolves its full dependency tree (deduped, cycle-safe) and installs every required skill alongside it. Use --no-deps to install just the meta skill.

npx @dmzagent/skill-builder install @skillauthor/my-bundle
# → resolves and installs my-bundle + every skill it depends on

Getting started: publishing skills

1. Create an account

npx @dmzagent/skill-builder register myhandle "My Name" me@email.com

You'll be prompted for a password. Or set SKILL_PASSWORD to skip the prompt:

SKILL_PASSWORD=hunter2 npx @dmzagent/skill-builder register myhandle "My Name" me@email.com

2. Log in

npx @dmzagent/skill-builder login me@email.com

Your auth token is printed on success. Set it as SKILL_TOKEN so subsequent commands can use it:

export SKILL_TOKEN=<token-from-login>

(Or pass --token $SKILL_TOKEN to each command.)

3. Publish a skill

From a markdown file:

npx @dmzagent/skill-builder publish ./my-skill.md

The CLI reads the filename as the skill id and the first heading as the name. You can also use a JSON manifest:

npx @dmzagent/skill-builder publish ./my-skill.json

4. Declare the capabilities a skill needs (optional)

A skill can declare what a model must be able to do to invoke it. The declaration lives on the spec, so it travels with the skill through the registry, the CLI, and MCP — an agent can tell whether it is able to run a skill before it spends a turn on it.

{
  "capabilities": [
    { "id": "vision", "level": "required", "note": "reads chart images" },
    { "id": "long-context", "level": "preferred" }
  ]
}

required capabilities gate execution: the registry refuses to run a skill on a runtime that cannot provide one, unless the caller explicitly forces a degraded run. preferred capabilities only weaken the result. Catalog ids:

vision · audio-input · file-input · structured-output · streaming
extended-reasoning · long-context · multilingual
tool-use · parallel-tool-calls · code-execution · web-search · file-system
computer-use · mcp-client
persistent-memory · citations

Custom ids are accepted for anything the catalog does not model yet. info and list show the contract, and install rolls it up across a meta skill's whole dependency tree — a meta skill only runs where every skill it installs can run.

npx @dmzagent/skill-builder info @skillauthor/chart-reader
# Required capabilities (1) — the invoking model must provide these:
#   - vision (required) — reads chart images

5. Fork an existing skill (optional)

npx @dmzagent/skill-builder fork @skillauthor/dialogue-flow --name my-dialogue

Creates a copy under your account that you can republish.

Maintaining the registry: backfill dependencies

scripts/backfill-dependencies.mjs finds existing skills that implicitly orchestrate other skills and proposes explicit dependencies + type: "meta" for them. It reads every skill, detects references to other real registry skills (by full id, install directives, or distinctive names), scores a confidence level, and merges a curated set of known-correct mappings.

It is dry-run by default — it prints a report and writes nothing:

node scripts/backfill-dependencies.mjs                 # report only
node scripts/backfill-dependencies.mjs --json          # machine-readable

To apply, pass --apply with a token. Only skills owned by the token's account are updated (others are listed and skipped), so run it once per publishing account:

SKILL_TOKEN=<token> node scripts/backfill-dependencies.mjs --apply
node scripts/backfill-dependencies.mjs --apply --include-medium   # also apply medium-confidence

Confidence tiers: curated and high (explicit id / install / distinctive-name references) apply by default; medium (fuzzy name match) only with --include-medium; meta-no-deps (orchestration language but no concrete dependency) is reported for manual review and never auto-applied.

Applying requires the registry to run the worker build that understands type/dependencies (it derives type and regenerates markdown on save). The dry-run report works against any deployed version.

Use from coding agents (MCP)

Coding agents — Claude Code, Cursor, Codex, and other MCP clients — can search the registry and auto-install skills (resolving meta-skill dependency trees) via the @dmzagent/skill-builder-mcp MCP server.

// e.g. .mcp.json (Claude Code) or .cursor/mcp.json (Cursor)
{
  "mcpServers": {
    "skill-builder": { "command": "npx", "args": ["-y", "@dmzagent/skill-builder-mcp"] }
  }
}

Tools: skill_search, skill_info, skill_suggest, skill_taxonomy, and skill_install (writes skill files for the detected tool). See packages/mcp/README.md for per-client setup.

Web App

Visit the skill studio at https://skills.dmzagent.com. The workspace is three collapsible panes plus a settings drawer:

• Skill Architect — an agent that edits the spec directly rather than handing back markdown, with suggested next steps drawn from whatever the spec is still missing.
• Spec canvas — the skill as a document: capability contract, instructions, prompt template with its {{variables}} highlighted, examples and tests, and a publish-readiness checklist that links into the field behind each gap.
• Preview — run the skill for real. Template variables become fields, a capability preflight says whether the runtime can honestly execute it, and every run is scored against the expectation it came from. Any run can be handed back to the architect for review.
• Architecture — for meta skills, the dependency tree as a layered graph with unresolved ids and cycles flagged, and the capability contract rolled up across everything the skill installs.
• Settings drawer — every field of the spec, out of the way until you want it: identity, behavior, capabilities, composition, examples and tests, and the generated markdown artifact.

Install from source

npm install
npm run dev --workspace packages/web

License

MIT