diff --git a/docs/phone-first-setup.md b/docs/phone-first-setup.md
new file mode 100644
index 0000000..bff6971
--- /dev/null
+++ b/docs/phone-first-setup.md
@@ -0,0 +1,74 @@
+# Phone-first setup: drive your agents from anywhere
+
+This guide is for new users who want the full remote workflow: every coding
+agent reachable from a phone, one team room for coordination, and approval
+buttons for anything gated. It assumes you have at least one always-on or
+frequently-on machine (desktop, mini, laptop) where an agent runs.
+
+## The mental model: three surfaces
+
+Getting this straight up front saves hours of confusion:
+
+1. **The agent's session** — its actual working brain. One conversation,
+   living on the agent's machine. The provider app (Claude app for Claude
+   Code agents, ChatGPT app for Codex agents) can open a **window into this
+   same session**: what you type on the phone lands in the session, what the
+   agent does shows on the phone. Same conversation, two screens.
+2. **The team room** — a shared channel (GroupMind) that you and every agent
+   read and post to. Agents wake within seconds of a new message. This is
+   where coordination happens: task handoffs, decisions, status.
+3. **CodeWatch** — the phone app that shows the room plus **Approve/Deny
+   buttons** for gated actions (merges, privileged commands). Approvals are
+   one tap; agents verify the button decision server-side.
+
+Rule of thumb: **talk to the team in the room; type into an agent's session
+only when you need to drive that specific agent directly.**
+
+## One session per agent
+
+Each agent has exactly **one** canonical session, ever. Don't open a second
+chat for a side task — you end up with two half-informed copies of the same
+agent, and (before ide-agent-kit's responder lock) both would answer the
+team room as the same handle with conflicting messages.
+
+ide-agent-kit enforces the room half of this automatically: the SessionStart
+bootstrap claims a per-machine **responder lock**, so if a stray session
+starts anyway, it boots passive — it serves you directly if you type into
+it, but it never arms the room loop or posts as the agent. The first session
+keeps the room voice; the lock releases itself when that session's process
+exits.
+
+## Setup, per agent machine
+
+1. **Install ide-agent-kit** and run `ide-agent-kit init --ide claude-code`
+   in the agent's working directory. This installs two hooks: the room
+   poller (message notifications) and the session bootstrap (instant wake +
+   responder lock). No manual arming — a fresh session self-arms.
+2. **Connect the agent to the team room.** In CodeWatch: add the room, then
+   add the agent with its own handle and login key (each agent gets its own
+   handle — don't share one key across agents).
+3. **Make the session phone-drivable.** Two paths, either works:
+   - Start the agent's session **from the provider app** (Claude app →
+     Code tab → new session on that machine). App-created sessions are
+     drivable by construction.
+   - Or, in an **interactive** terminal session, run `/remote-control` —
+     it registers the running session, full history included. (This command
+     is unavailable in headless/automation-launched sessions; that's
+     expected — use the app-created path instead.)
+4. **Verify.** Type a marker word into the phone entry; the agent should see
+   it in-session within seconds. Post a message in the team room; the agent
+   should react without being poked.
+
+## What to expect day-to-day
+
+- **From your phone you can:** follow and steer everything in the room,
+  approve gated actions with one tap in CodeWatch, and drop into any
+  agent's session directly via the Claude/ChatGPT app when it needs
+  hands-on driving.
+- **Laptop agents sleep.** An agent on a laptop is unreachable while the
+  lid is closed or the machine suspends. Put agents that must always be
+  reachable on an always-on machine.
+- **Stale session entries.** If you kill an agent's process, its entry may
+  linger in the provider app's session list. It's inert — archive it from
+  the app. Content matching (does the entry show the agent's real work?) is
+  the reliable way to identify the live session.