1. Docs
2. Reference

Make your agents great at this

Arcon works with any agent as-is. But an agent that knows you’re following from a phone is much nicer to follow: a clear status line, a real checklist, questions you can answer with a tap, and documents instead of walls of text. That’s what the Arcon skill teaches.

5 min read

What the Arcon skill does

A skill is a set of instructions an agent loads when it’s relevant. arcon install skill (also run by arcon install claude-code) puts the Arcon skill in ~/.claude/skills/arcon for Claude Code and adds the same rules to ~/.codex/AGENTS.md for Codex, between Arcon’s own markers, leaving the rest of the file alone.

Claude Code picks the skill up whenever Arcon is paired, or when you mention your phone, approvals or “keep me posted”. Codex reads AGENTS.md at the start of every session. After updating Arcon, run arcon install skill again to get the newest version.

The loop your agent follows

Every task, big or small, goes through six moves. Here’s what each one looks like from your phone:

1. Frame

   One line you’d read on a lock screen: “Migrating 14 endpoints to v2”.

   On your phone: The session’s status line
2. Plan

   A checklist of 3 to 8 outcomes, not activities.

   On your phone: The checklist and progress bar
3. Execute

   Quiet work; a heads-up before anything irreversible.

   On your phone: Ticks, and the occasional one-liner
4. Ask

   Only when blocked: one question, 2–4 options, a suggested default.

   On your phone: One notification, answered with a tap
5. Publish

   Anything worth reading becomes a Markdown document.

   On your phone: A clean document in Docs
6. Close

   A five-line summary: result, what changed, how to verify.

   On your phone: “Done — …” in the chat

The skill also keeps agents frugal (plan first, read only what’s needed, don’t narrate) and careful: reads just happen, hard-to-undo changes get a heads-up, and anything that sends, deploys or pays gets a question first, even when the agent technically has permission.

A good close looks like this

Done — checkout p95 is down from 4.1 s to 0.3 s.Changed: src/api/cart/price.ts, one batched query instead of 40.Verify: pnpm test --filter cart && pnpm bench checkoutNote: the old cache flag is now unused; left in place.Details: "Checkout latency — root cause & fix" on your phone.

How to prompt for the phone

The skill does most of the work. These habits do the rest:

• Say you’re leaving. “I’m heading out, keep me posted on my phone” makes the agent turn on Take over, so the conversation continues in the app.
• Give the goal and what “done” means. The checklist can only be as clear as your request.
• Pre-decide what you can, and say which decisions you want to make yourself: “ask me before choosing a library”.
• Ask for one question at a time, with options. Tapping beats typing on a phone.
• Ask for a document when you want to read something: “publish a short plan first, then wait for my go”.
• Draw the lines: “don’t push, deploy or delete anything without asking”.

The commands agents use

The skill uses a handful of arcon commands. You rarely type them, but it helps to know what they do. When Arcon isn’t running they do nothing and never fail, so they’re safe in any script.

Every flag is in the CLI reference.