diff --git a/README.md b/README.md
index 095d2c0..fcd311e 100644
--- a/README.md
+++ b/README.md
@@ -8,6 +8,8 @@ Reusable OpenClaw skill for `/auto-continue` (`/ac`) session-level auto-followup
 
 ### 這個 repo 內有什麼
 - `openclaw-auto-continue/SKILL.md` — **給 Agent 看的** skill 指引
+- `openclaw-auto-continue/references/installation-guide.md` — **給要落地 `/ac` 指令的人 / agent 看** 的安裝指南
+- `openclaw-auto-continue/references/runtime-integration.md` — `/ac` runtime integration 說明
 - `openclaw-auto-continue.skill` — 已打包好的 skill 檔
 
 ### 其他電腦怎麼用
@@ -97,7 +99,8 @@ cp -R openclaw-auto-continue ~/.openclaw/workspace/skills/
 一句話：
 - `README.md` 是給人 / operator 看的
 - `openclaw-auto-continue/SKILL.md` 是給 agent 看的
-- `references/runtime-integration.md` 是給要落地 runtime 的人 / agent 看的
+- `references/installation-guide.md` 是給要從頭做到尾落地 `/ac` 的人 / agent 看的
+- `references/runtime-integration.md` 是給要理解 runtime 接法的人 / agent 看的
 
 ---
 
@@ -105,6 +108,8 @@ cp -R openclaw-auto-continue ~/.openclaw/workspace/skills/
 
 ### What this repo contains
 - `openclaw-auto-continue/SKILL.md` — the **agent-facing** skill instructions
+- `openclaw-auto-continue/references/installation-guide.md` — end-to-end guide for making `/ac` real on another machine
+- `openclaw-auto-continue/references/runtime-integration.md` — runtime integration notes for `/ac`
 - `openclaw-auto-continue.skill` — packaged skill artifact
 
 ### For humans: how to use on another computer
@@ -194,4 +199,5 @@ If an OpenClaw agent is deciding how to use this repo:
 In short:
 - `README.md` is the human/operator guide
 - `openclaw-auto-continue/SKILL.md` is the agent-facing guide
-- `references/runtime-integration.md` is for runtime integration work
+- `references/installation-guide.md` is for end-to-end `/ac` installation work
+- `references/runtime-integration.md` is for runtime integration details
diff --git a/openclaw-auto-continue/SKILL.md b/openclaw-auto-continue/SKILL.md
index 76698a7..5bc4fc1 100644
--- a/openclaw-auto-continue/SKILL.md
+++ b/openclaw-auto-continue/SKILL.md
@@ -124,7 +124,7 @@ A practical verification flow:
 3. Give the agent a multi-step but low-risk task
 4. Confirm the agent continues without waiting after every intermediate step
 
-If the target machine does not recognize `/ac`, read `references/runtime-integration.md` and treat it as a runtime integration problem, not a skill-discovery problem.
+If the target machine does not recognize `/ac`, read `references/installation-guide.md` first, then `references/runtime-integration.md` if deeper implementation detail is needed. Treat it as a runtime integration problem, not a skill-discovery problem.
 
 If you need concrete implementation evidence, the verified live installation path was:
 - `/home/chchang/.nvm/versions/node/v22.22.0/lib/node_modules/openclaw/dist/commands-handlers.runtime-Akj_Dqoi.js`
diff --git a/openclaw-auto-continue/references/installation-guide.md b/openclaw-auto-continue/references/installation-guide.md
new file mode 100644
index 0000000..f7c49f9
--- /dev/null
+++ b/openclaw-auto-continue/references/installation-guide.md
@@ -0,0 +1,285 @@
+# Installation Guide: build `/auto-continue` (`/ac`) into a target OpenClaw runtime
+
+This guide is written so an agent can follow it end-to-end and add `/auto-continue` (`/ac`) support to another OpenClaw machine.
+
+Use this when the target machine has the skill/docs but the live runtime does not actually recognize `/ac`.
+
+## Goal
+
+After following this guide, the target machine should support:
+
+- `/ac status`
+- `/ac on`
+- `/ac off`
+- `/auto-continue status`
+- `/auto-continue on`
+- `/auto-continue off`
+
+with session-level behavior equivalent to:
+
+- on:
+  - `queueMode = "followup"`
+  - `queueDebounceMs = 0`
+  - `queueCap = 1`
+  - `queueDrop = "summarize"`
+- off:
+  - remove those session queue fields
+  - clear queued followup/lane items for that session
+
+---
+
+## Phase 0 — Preconditions
+
+Before editing code, confirm all of these:
+
+1. You can access the target machine's OpenClaw installation.
+2. You know which installation is actually running.
+3. You can restart the OpenClaw service/process after changes.
+4. You can test from a real chat surface after deployment.
+
+If any of these are unknown, stop and resolve them first.
+
+---
+
+## Phase 1 — Locate the live runtime
+
+Do **not** assume the checked-out repo is the runtime.
+
+Find the actual OpenClaw installation used by the running service. On the verified machine, the live handler existed at:
+
+- `/home/chchang/.nvm/versions/node/v22.22.0/lib/node_modules/openclaw/dist/commands-handlers.runtime-Akj_Dqoi.js`
+
+What matters is not this exact path, but the principle:
+
+- identify the real running install
+- patch the command handler in the live runtime or rebuild/redeploy the source that produces it
+
+### Verification checklist
+
+Before patching, confirm:
+- which binary/service launches OpenClaw
+- which code/dist that service is using
+- whether slash commands are already handled in that runtime
+
+---
+
+## Phase 2 — Find the command handler entry point
+
+Search the target runtime for the existing slash-command dispatch logic.
+
+You need the place where incoming messages/commands are parsed and existing commands are handled.
+
+Typical target shape:
+- a function that receives inbound text
+- a branch that recognizes slash commands
+- session lookup/load logic
+- persistence/save logic for session state
+
+If the target runtime already has command handlers, add `/auto-continue` and `/ac` there.
+
+---
+
+## Phase 3 — Add `/ac` command recognition
+
+Add recognition for both:
+
+- `/auto-continue`
+- `/ac`
+
+and their argument forms:
+
+- `/ac on`
+- `/ac off`
+- `/ac status`
+- `/auto-continue on`
+- `/auto-continue off`
+- `/auto-continue status`
+
+### Parse rule
+
+Normalize input into one of:
+- `status`
+- `on`
+- `off`
+
+Recommended default:
+- if no argument is supplied, treat it as `status`
+
+---
+
+## Phase 4 — Add handler behavior
+
+Implement behavior equivalent to the following pseudocode:
+
+```js
+if (command is /ac or /auto-continue) {
+  authorizeSender();
+
+  const action = parseAction(arg); // status | on | off
+  const sessionEntry = await loadSessionEntry(sessionKey);
+
+  if (action === 'status') {
+    return replyWithCurrentQueueMode(sessionEntry);
+  }
+
+  if (action === 'on') {
+    sessionEntry.queueMode = 'followup';
+    sessionEntry.queueDebounceMs = 0;
+    sessionEntry.queueCap = 1;
+    sessionEntry.queueDrop = 'summarize';
+    await saveSessionEntry(sessionEntry);
+    return replyEnabled();
+  }
+
+  if (action === 'off') {
+    delete sessionEntry.queueMode;
+    delete sessionEntry.queueDebounceMs;
+    delete sessionEntry.queueCap;
+    delete sessionEntry.queueDrop;
+    await saveSessionEntry(sessionEntry);
+    await clearQueuedItemsForSession(sessionKey);
+    return replyDisabled();
+  }
+}
+```
+
+### Required semantics
+
+#### `status`
+Return enough information to confirm whether the session is:
+- default / unset, or
+- using `followup`
+
+#### `on`
+Must persist session settings so the behavior survives beyond the immediate in-memory branch.
+
+#### `off`
+Must:
+- remove the session-level queue config
+- clear queued followup/lane items for that session so old queued actions do not keep running
+
+---
+
+## Phase 5 — Preserve safety boundaries
+
+`/ac` must **not** bypass:
+- approvals
+- safety checks
+- destructive action confirmation
+- decision gates that still require the user
+
+This is a queue/followup behavior switch, not unrestricted autonomy.
+
+---
+
+## Phase 6 — Deploy the change
+
+Use the deployment path appropriate for the target machine:
+
+### Option A — patch live dist directly
+Use this only when you intentionally maintain a live dist patch.
+
+### Option B — patch source, then rebuild/redeploy
+Preferred when the target machine is managed from source.
+
+After deployment:
+1. restart the OpenClaw process/service
+2. confirm the restarted service is using the new runtime
+3. verify the target chat surface still routes commands normally
+
+---
+
+## Phase 7 — Verify on a real chat
+
+Use a real chat/session on the target machine.
+
+### Test 1 — command recognition
+Run:
+
+```text
+/ac status
+```
+
+Expected:
+- no unknown-command failure
+- a status reply describing the current queue mode
+
+### Test 2 — enable
+Run:
+
+```text
+/ac on
+/ac status
+```
+
+Expected:
+- status now shows followup mode or equivalent enabled state
+
+### Test 3 — practical behavior
+Give a low-risk, multi-step task.
+
+Expected:
+- the agent is more likely to continue into the next step
+- the agent pauses less between intermediate steps
+- approvals/safety still stop it when appropriate
+
+### Test 4 — disable
+Run:
+
+```text
+/ac off
+/ac status
+```
+
+Expected:
+- status returns to default/unset behavior
+- any queued followup work for that session is cleared
+
+---
+
+## Phase 8 — Troubleshooting
+
+### Problem: `/ac` is unknown
+Likely causes:
+- you edited the wrong installation
+- service restarted into a different runtime
+- the surface does not pass through slash commands
+- the handler was added in source but not rebuilt/deployed
+
+### Problem: `/ac status` works but `/ac on` has no effect
+Check:
+- whether the session entry is actually being saved
+- whether you are testing the same session you enabled
+- whether queue fields are being overwritten elsewhere
+
+### Problem: `/ac off` replies successfully but the agent still continues
+Check:
+- whether queued followup/lane items were really cleared
+- whether a different queue path exists in this runtime
+- whether the next message was already in flight before disable completed
+
+### Problem: it still pauses sometimes even when enabled
+That can be normal if the next step is blocked by:
+- approval requirements
+- safety checks
+- destructive actions
+- explicit user choice points
+
+---
+
+## Acceptance criteria
+
+Do not claim completion unless all are true:
+
+1. `/ac status` is recognized on the target machine
+2. `/ac on` changes the session to followup mode
+3. `/ac off` removes that mode
+4. queued followup items are cleared on disable
+5. a real chat test confirms improved auto-continue behavior
+6. approvals/safety behavior is unchanged
+
+---
+
+## One-line rule
+
+**The skill teaches usage, but this installation guide is the procedure that makes `/ac` real on another machine.**