Agent skills

The OriginAI skill teaches a coding agent how to use the CLI, treat releases as the source of truth, and author and validate RPML. Installing it is what turns a generic agent into one that understands your OriginAI workflow.

What the skill does

• Connects Claude Code, Cursor, Codex, Pi, Hermes, and skills.sh agents to your OriginAI product specs.
• Tells the agent to use MCP tools first when connected, and to fall back to the originai CLI (which has the same semantics) otherwise.
• Documents the release-default read model and the post-release change-request write loop.
• Bundles RPML references under rpml/ — the method, an element index, examples, and prompts.
• Relies on npx originai login for auth rather than embedding tokens in the skill.

Install the skill

Option A — skills.sh (the recommended discovery path)

The public skill lives in 21stware/originai-context under skills/origin-product-spec-management/.

npx skills add 21stware/originai-context
# or pin a specific skill:
npx skills add 21stware/originai-context --skill origin-product-spec-management
# or list what's available first:
npx skills add 21stware/originai-context --list

The skill contains no secrets and no project ID.

• Existing Origin project: npx originai login, then npx originai link --project <id>.
• No project yet (repo → Origin): install the skill, log in, and paste the prompt from Create product specs from a codebase. The agent runs create-project. You do not need link --project first.

You can also install from the site's well-known endpoint (the same skill body over HTTP):

npx skills add https://getoriginai.com
# https://getoriginai.com/.well-known/agent-skills/index.json

For more, see skills.sh and the skills CLI.

Option B — the OriginAI CLI (writes the skill into your repo)

npx originai login
npx originai link --project <project-id> --skill
# or install the skill only, if .origin.json already exists:
npx originai link --skill

Option C — the Claude Code plugin

Install the Origin marketplace plugin from the same public repo; it bundles the skill and a local MCP server. See MCP & agent install.

After installing, ask the agent to read SKILL.md and either implement a published release or write specs from the repo.

Start from a codebase

Prefer Claude Code or Codex with a frontier model. Install the skill (above), then paste the prompt from the create page (Write specs from the repo) or from Create product specs from a codebase. The agent creates the Origin project and writes specs. You review, then publish.

Setup, end to end

1. Install the skill — via npx skills add 21stware/originai-context, originai link --skill, or the Claude plugin.
2. Authorize the CLI — npx originai login writes ~/.origin/settings.json.
3. Bind or create the project — npx originai link --project <id> writes .origin.json. From an unbound repo, the agent can create-project instead.
4. Let the agent work — through MCP tools when connected, or the originai CLI otherwise.

How sync works

• Read commands use the latest published release, so agents implement formally released specs rather than in-progress workspace edits.
• Write commands modify the live workspace before the first release. After that, they stage a change request (suggested: true) for you to apply, then publish. Agents never apply their own requests.
• .origin.json carries the release_hash. After implementing a published release, originai sync advances the pointer.

For the full command table, see the OriginAI CLI.

HTTP discovery (optional)

The CLI is preferred, but an agent can also fetch the skill over HTTP without installing anything:

• https://getoriginai.com/.well-known/agent-skills/index.json
• https://getoriginai.com/skills-full.txt

These files are generated from SKILL_MD at frontend build time (frontend/scripts/generate-well-known-skill.ts). Only ever edit SKILL_MD.

Keeping the skill accurate

Related

• OriginAI CLI
• Connect a coding agent
• Workspace vs release
• Review change requests