Operon — Setup

A meta-skill: install the Operon CLI on a fresh machine, authenticate the user via a browser tab, detect their AI tools, pick or create a workspace, mine their first candidates, and surface the top three so they can publish one. Designed to be driven by an AI assistant (Claude Code, Codex, Cursor) end-to-end.

When to use this skill

• The user says "set up Operon for me" or "install Operon."
• The user has a recurring prompt they keep retyping in Claude Code,

Codex, or Cursor and asks how to share it with a teammate.

• The user wants to see what AI workflows recur across their sessions

and doesn't know where to start.

How to use it

1. Check prerequisites

node --version

Require node ≥ 22. If older, point them at https://nodejs.org and stop.

2. Install the CLI

npm i -g @withoperon/cli

If they're on a system without global npm permissions, suggest the pnpm dlx @withoperon/cli form instead.

3. Run the unified setup verb

operon setup --yes --json

Three important things about this invocation:

• `--yes` accepts every default that has one. Workspace name will

default to the user's $USER; mining window defaults to 30d.

• `--json` switches the CLI into NDJSON mode — one JSON object per

line is written to stdout the instant each step completes. Parse stdout line by line instead of waiting for one final blob.

• The CLI opens the user's browser to a URL like

https://app.withoperon.com/auth/cli?code=ABC12345. They click Approve. No copy/paste of tokens.

4. Parse NDJSON events as they land

Each event has the shape { ts, cmd, level, event, data }. Watch for:

If the user closes the terminal mid-setup, you can recover state with:

cat ~/.operon/events.log | tail -50

5. Present the top candidates

After mine.complete, you have a list like:

{
  "slug": "review-pr",
  "score": 0.91,
  "uses": 12,
  "summary": "Review the diff for security, performance, and consistency issues."
}

Show the user the top 3 in plain English and ask which to publish.

6. (Optional) Polish the chosen candidate with Claude

Only if ANTHROPIC_API_KEY is set in the environment:

operon polish <slug>

This drafts a clean SKILL.md from the raw candidate. Skip if the key isn't set — operon publish works without polishing.

7. Publish

operon publish <slug>

Report the resulting URL back to the user (it'll look like https://app.withoperon.com/<workspace>/<slug>).

Error handling

If a command exits non-zero, parse stderr for the line https://withoperon.com/docs/cli/errors#<slug>. The error code prefix tells you the recovery step:

• `AUTH_REQUIRED` → re-run operon setup. Their token is missing

or invalid.

• `NO_SOURCES` → tell the user to install Claude Code, Codex, or

Cursor first. Operon needs at least one to mine.

• `NO_CANDIDATES` → retry with --since 90d or --since 180d.

Their sessions don't have enough repeats yet in the current window.

• `WORKSPACE_NOT_LINKED` → re-run operon setup (or pass

--workspace <slug> if they know the slug).

• `RATE_LIMITED` → wait a minute and retry.
• `NETWORK_ERROR` → check connectivity; if they self-host, pass

--api-base <url>.

• `USER_ABORT` → exit code 0; they canceled. Re-run when ready.

Strict (CI / non-interactive) mode

If you're driving the CLI in a fully-automated context where no human will see prompts, also pass --non-interactive. The CLI will fail fast with a clear error code if any input was needed instead of hanging.

operon setup --yes --json --non-interactive

What setup does NOT do

• Doesn't install Claude Code, Codex, or Cursor for the user. Those are

prerequisites; if none are present, setup aborts with NO_SOURCES.

• Doesn't auto-publish anything. The mine step is --dry-run — the

user explicitly opts in via operon publish.

• Doesn't read session content beyond what Operon itself needs to

cluster recurring prompts. All mining is local-first; nothing leaves the laptop until publish.

Notes for agents

• The CLI writes its config to ~/.operon/config.json. After the

first successful operon setup, subsequent runs are authenticated automatically — you don't need to repeat the device-code dance.

• If you see a started event but no auth.ready after ~30 seconds,

the user hasn't clicked Approve yet — surface the URL from auth.browser_opened.data.url so they can find the tab.

• The setup verb is idempotent. Calling it twice does the right thing:

step 1 short-circuits if already authed, step 3 short-circuits if a workspace is already linked, step 4 short-circuits if candidates already exist on disk.