diff --git a/docs/parallel_training/cluster.rst b/docs/design.rst
similarity index 89%
rename from docs/parallel_training/cluster.rst
rename to docs/design.rst
index ce688ce2e912621e89329c10c6032371545cad96..b2a76a898707b86d890671d8dae9dedfbd07b862 100644
--- a/docs/parallel_training/cluster.rst
+++ b/docs/design.rst
@@ -1,11 +1,8 @@
 Parl Cluster
 ============
 
-Get Started
-###########
-
 Cluster Structure Overview
---------------------------
+##########################
 
 | There are three core concepts in a Parl cluster: master, worker and client.
 
@@ -19,12 +16,13 @@ Cluster Structure Overview
 - **Client:** For each training program, there is a unique global client which
   submits tasks to the master node.
 
-.. image:: ./cluster_structure.png
+.. image:: ./parallel_training/cluster_structure.png
   :width: 600px
   :align: center
 
 Master
-------
+######
+
 | There is only one master node in each parl cluster, we can start a master by
   calling ``xparl start --port 1234`` with a assigned port number. This command
   will also simultaneously start a local worker which connects to the new
@@ -43,14 +41,14 @@ Master
   status of the cluster (i.e. total cpu number, used cpu number, load average
   ...) to the monitor.
 
-.. image:: ./master.png
+.. image:: ./parallel_training/master.png
   :width: 600px
   :align: center
 
 Worker
-------
+######
 
-| We can add more computation resources to a existed cluster by calling
+| We can add more computation resources to an existed cluster by calling
   ``xparl --connect master_address`` command. This command will create a local
   **Worker** object and then connect to the cluster.
 
@@ -62,27 +60,27 @@ Worker
   job from the job buffer, start a new job and update worker information to
   the master node.
 
-.. image:: ./worker.png
+.. image:: ./parallel_training/worker.png
   :width: 600px
   :align: center
 
 Client
-------
+######
 
 | We have a global client for each training program, it submits training tasks
   to the master node. User do not need to interact with client object directly.
-  We can create a new global client or get an existed global client by calling
+  We can create a new global client and connect it to the cluster by calling
   ``parl.connect(master_address)``.
 
 | The global client will read local python scripts and configuration files,
   which will later be sent to remote jobs.
 
-.. image:: ./client.png
+.. image:: ./parallel_training/client.png
   :width: 600px
   :align: center
 
 Actor
------
+#####
 
 | **Actor** is an object defined by users which aims to solve a specific task.
   We use ``@parl.remote_class`` decorator to convert an actor to a
@@ -108,6 +106,6 @@ Actor
 | When the actor call a function, the real computation will be executed in the
   job process by job's local actor.
 
-.. image:: ./actor.png
+.. image:: ./parallel_training/actor.png
   :width: 600px
   :align: center
diff --git a/docs/images/parl.graffle b/docs/images/parl.graffle
index 021d34a2b2e3381e97379ec4fb81e6832ddf0037..b0b0d57b3465a8836be03b87ecd8b086a3e7d590 100644
Binary files a/docs/images/parl.graffle and b/docs/images/parl.graffle differ
diff --git a/docs/index.rst b/docs/index.rst
index 9e92ec79d9b499315ded8026891e8d60c988b0b2..3235aec000f3c817a2de1689786b02d67250d280 100644
--- a/docs/index.rst
+++ b/docs/index.rst
@@ -66,7 +66,6 @@ Abstractions
     :caption: Parallel Training
 
     parallel_training/overview.rst
-    parallel_training/cluster.rst
     parallel_training/setup.rst
     parallel_training/recommended_practice.rst
 
@@ -75,6 +74,7 @@ Abstractions
    :caption: High-quality Implementations
 
    implementations.rst
+   design.rst
 
 .. toctree::
    :maxdepth: 1
diff --git a/docs/parallel_training/client.png b/docs/parallel_training/client.png
index 5d06abe4498c8cd8225296479674616b887f5603..688538310447f9701e5132f72880ac2c7baf624b 100644
Binary files a/docs/parallel_training/client.png and b/docs/parallel_training/client.png differ
diff --git a/docs/parallel_training/cluster_structure.png b/docs/parallel_training/cluster_structure.png
index 3c24a2d721b8afed374e901b4b36cd63e83d2bb1..ce4829bc44f8e1656ed5487d7e35c4263a2ef76e 100644
Binary files a/docs/parallel_training/cluster_structure.png and b/docs/parallel_training/cluster_structure.png differ
diff --git a/docs/parallel_training/master.png b/docs/parallel_training/master.png
index 21d51cabac641c4e9a24fa410fbd33b1637f933a..093da84a55d6d1c2806f5c7094bc2c21d48dc21d 100644
Binary files a/docs/parallel_training/master.png and b/docs/parallel_training/master.png differ
diff --git a/docs/parallel_training/worker.png b/docs/parallel_training/worker.png
index 93944ec5b98d45a16b225de534caaa88bfcc0a81..ebe2a9bbb8e3232843315f51bd8e1f2f36ae66ad 100644
Binary files a/docs/parallel_training/worker.png and b/docs/parallel_training/worker.png differ
diff --git a/parl/remote/client.py b/parl/remote/client.py
index 9b20e404c237896a7f8c0cc2d9ec45910edf17f9..eb09b109812c91d9cb1b5b11c3190e3be099da72 100644
--- a/parl/remote/client.py
+++ b/parl/remote/client.py
@@ -32,8 +32,8 @@ class Client(object):
     connect to the same global client in a training task.
 
     Attributes:
-        submit_task_socket (zmq.Context.socket): A socket which submits job to
-                                                the master node.
+        submit_task_socket (zmq.Context.socket): A socket which submits tasks to
+                                                 the master node.
         pyfiles (bytes): A serialized dictionary containing the code of python
                          files in local working directory.
         executable_path (str): File path of the executable python script.