diff --git a/src/interface.jl b/src/interface.jl
index ee8594c..59e0596 100755
--- a/src/interface.jl
+++ b/src/interface.jl
@@ -120,19 +120,21 @@ function chat_mistral_openorca(a::agentReflex)
         "
   """
 
+  conversation = messagesToString(a.messages)
+
   prompt = 
   """
-    <|im_start|>system
+    <|system|>
     $(a.roles[a.role])
     Your earlier talk with the user:
     $(a.earlierConversation)
-    <|im_end|>
-    $(messagesToString(a.messages))
+    <|/s|>
+    $conversation
     <|im_start|>assistant
-
   """
   
   response = sendReceivePrompt(a, prompt)
+  response = split(response, "\n\n")[1]
   response = split(response, "<|im_end|>")[1]
 
   return response
@@ -181,7 +183,7 @@ function planner_mistral_openorca(a::agentReflex)
   
   assistant_plan_prompt =
     """
-    <|im_start|>system
+    <|system|>
     $(a.roles[a.role])
     The required info you need for wine recommendation:
         - type of food: ask the user
@@ -190,41 +192,115 @@ function planner_mistral_openorca(a::agentReflex)
         - wine price range: ask the user
         - ambient temperature at the serving location: ask the user
         - wines we have in stock
-    You provide a personalized recommendation of wine based on the user's info above by describing the benefits of each wine in detail.
     
     You have access to the following tools:
     $toollines
 
-    Your work:
+    Your earlier work:
     $shorttermMemory
 
-    Use the following format:
-    Objective: the objective you intend to do
-    Aware: according to your work and your conversation with the user, ask yourself what info you have and what info you doesn't have?
-    Plan: first you should always think about the objective, the info you have, the info you doesn't have thoroughly then extract and devise a complete, step by step plan (pay attention to correct numeral calculation and commonsense).
+    Your task is to do the following:
+    Plan: first you should always think about your conversation with the user and your earlier work thoroughly then extract and devise a complete, step by step plan to achieve your objective (pay attention to correct numeral calculation and commonsense).
     P.S.1 each step of the plan should be a single action.
     P.S.2 ask the user if you don't have info.
-
-    <|im_end|>
+    <|/s|>
     $conversation
-    <|im_start|>assistant
-    Objective: 
+    <|assistant|>
+    Plan: 
     """
-  #WORKING remove () in steps as LLM sometimes use for (addtional info) but interfere with updatePlan
-  
 
-  result = sendReceivePrompt(a, assistant_plan_prompt, max_tokens=512, temperature=0.1)
+  plan = sendReceivePrompt(a, assistant_plan_prompt, max_tokens=512, temperature=0.1)
+  plan = split(plan, "<|user|>")[1]
+  plan = split(plan, "<|assistant|>")[1]
+  plan = split(plan, "\n\n")[1]
 
-  x = split(result, "<|im_end|>")[1]
-  @show x
-  x = split(x, "Step")[1]
-  x = split(x, "Plan:")
-  objective = x[1]
-  plan = x[2]
-
-  return objective, plan
+  return plan
 end
 
+# function planner_mistral_openorca(a::agentReflex)
+#   """
+#     general prompt format:
+
+#     "
+#     <|im_start|>system
+#     {role}
+#     {tools}
+#     {thinkingFormat}
+#     <|im_end|>
+#     {context}
+#     <|im_start|>user
+#     {usermsg}
+#     <|im_end|>
+#     <|im_start|>assistant
+
+#     "
+
+#     Note:
+#       {context} = 
+#         "
+#         {earlierConversation}
+#         {env state}
+#         {shortterm memory}
+#         {longterm memory}
+#         "
+#   """
+
+#   conversation = messagesToString(a.messages)
+#   toollines = ""
+#   for (toolname, v) in a.tools
+#     if toolname ∉ [""]
+#       toolline = "$toolname: $(v[:description]) $(v[:input]) $(v[:output])\n"
+#       toollines *= toolline
+#     end
+#   end
+
+#   # skip objective and plan because LLM is going to generate new plan
+#   shorttermMemory = dictToString(a.memory[:shortterm], skiplist=["Objective:", "Plan 0:"])
+  
+#   assistant_plan_prompt =
+#     """
+#     <|system|>
+#     $(a.roles[a.role])
+#     The required info you need for wine recommendation:
+#         - type of food: ask the user
+#         - occasion: ask the user
+#         - user's personal taste of wine: ask the user
+#         - wine price range: ask the user
+#         - ambient temperature at the serving location: ask the user
+#         - wines we have in stock
+#     You provide a personalized recommendation of wine based on the user's info above by describing the benefits of each wine in detail.
+    
+#     You have access to the following tools:
+#     $toollines
+
+#     Your earlier work:
+#     $shorttermMemory
+
+#     Use the following format:
+#     Objective: what do you think the user needs?
+#     Plan: first you should always think about your conversation with the user and your earlier work thoroughly then extract and devise a complete, step by step plan (pay attention to correct numeral calculation and commonsense).
+#     P.S.1 each step of the plan should be a single action.
+#     P.S.2 ask the user if you don't have info.
+#     <|/s|>
+#     $conversation
+#     <|assistant|>
+#     Objective: 
+#     """
+
+
+#   result = sendReceivePrompt(a, assistant_plan_prompt, max_tokens=512, temperature=0.1)
+
+#   x = split(result, "<|user|>")[1]
+#   x = split(result, "<|assistant|>")[1]
+
+#   x = split(x, "Step")[1]
+#   x = split(x, "Plan:")
+#   objective = x[1]
+#   plan = x[2]
+
+#   return objective, plan
+# end
+
 """ Update the current plan.
 """
 function updatePlan(a::agentReflex)
@@ -242,7 +318,7 @@ function updatePlan(a::agentReflex)
 
   prompt =
   """
-    <|im_start|>system
+    <|system|>
     $(a.roles[a.role])
     The required info you need for wine recommendation:
         - wine price range: ask the user
@@ -266,13 +342,14 @@ function updatePlan(a::agentReflex)
     Plan: 1. Ask the user for their food type.
     Obs: It will be Thai dishes.
     Updated plan: 1. Ask the user for their food type (Thai dishes).
-    <|im_end|>
+    </s|>
     Updated plan: 
   """
 
   result = sendReceivePrompt(a, prompt, max_tokens=512, temperature=0.1)
   @show updatedPlan = result
   a.memory[:shortterm]["Plan 0:"] = result
+  
 end
 
 function actor_mistral_openorca(a::agentReflex)
@@ -323,13 +400,15 @@ function actor_mistral_openorca(a::agentReflex)
 
   prompt = 
   """
-    <|im_start|>system
+    <|system|>
     $(a.roles[a.role])
     You have access to the following tools:
     $toollines
-    $(a.thinkingFormat[:actor])
-    <|im_end|>
+    Your earlier work:
     $shorttermMemory
+    $(a.thinkingFormat[:actor])
+    <|/s|>
+    <|assistant|>
     Thought $(a.step): 
   """
   prompt = replace(prompt, "{toolnames}" => toolnames)
@@ -342,7 +421,7 @@ function actor_mistral_openorca(a::agentReflex)
   chunkedtext = nothing
 
   tempcounter = 0.0
-  while true  # while Thought is empty, run actor again
+  while true  # while Thought or Act is empty, run actor again
     tempcounter += 0.1
     @show tempcounter
     response = sendReceivePrompt(a, prompt, temperature=tempcounter)
@@ -375,8 +454,10 @@ function actor_mistral_openorca(a::agentReflex)
     chunkedtext = chunktext(response, headers) 
 
     # assuming length more than 10 character means LLM has valid thinking
-    if length(chunkedtext["Thought $(a.step):"]) > 10 
-      break
+    if haskey(chunkedtext, "Thought $(a.step):") && haskey(chunkedtext, "Act $(a.step):")
+      if length(chunkedtext["Thought $(a.step):"]) > 10 && length(chunkedtext["Act $(a.step):"]) > 10
+        break
+      end
     end
   end
   
@@ -483,12 +564,14 @@ end
 function work(a::agentReflex)
   workstate = nothing
   response = nothing
-
-  #BUG there is no Obs 2:
+  
   # user answering LLM -> Obs
-  if haskey(a.memory[:shortterm], "Act $(a.step):")
-    if occursin("chatbox", a.memory[:shortterm]["Act $(a.step):"])
-      a.memory[:shortterm]["Obs $(a.step):"] = a.messages[end][:content]
+  if length(a.memory[:shortterm]) != 0
+    latest_step = dictLatestStep(a.memory[:shortterm])
+    if haskey(a.memory[:shortterm], "Act $latest_step:")
+      if occursin("chatbox", a.memory[:shortterm]["Act $latest_step:"])
+        a.memory[:shortterm]["Obs $latest_step:"] = a.messages[end][:content]
+      end
     end
   end
 
@@ -499,41 +582,15 @@ function work(a::agentReflex)
     if a.attempt <= a.attemptlimit
       toolname = nothing
       toolinput = nothing
-
-      # if length(a.memory[:shortterm]) != 0
-      #   updatePlan(a)
-      #   @show updatedPlan = a.memory[:shortterm]["Plan 0:"]
-      # else
-      #   objective, plan = planner_mistral_openorca(a)
-      #   a.memory[:shortterm]["Objective:"] = objective
-      #   a.memory[:shortterm]["Plan $(a.attempt):"] = plan
-      #   a.memory[:log]["Plan $(a.attempt):"] = plan
-
-      #   println("")
-      #   @show objective
-      #   @show plan
-      # end
       
-      objective, plan = planner_mistral_openorca(a)
-      a.memory[:shortterm]["Objective:"] = objective
+      plan = planner_mistral_openorca(a)
       a.memory[:shortterm]["Plan $(a.attempt):"] = plan
       a.memory[:log]["Plan $(a.attempt):"] = plan
       a.step = 0  # reset because new plan is created
 
       println("")
-      @show objective
       @show plan
       
-
-      # sometimes LLM add not-need word I don't want
-      # plan = splittext(response, ["Step 1", "<|im_end|>", "Response", "Execution", 
-      #                           "Result", "Recommendation", "My response"])
-      # plan = replace(plan, "Plan:"=>"")
-      # println("")
-      # @show plan
-      
-      
-      
       println("")
       @show a.attempt
 
@@ -636,24 +693,25 @@ function actor(a::agentReflex)
   while true  # Actor loop
 
     # decide whether to repeat step or do the next step
-    decision, reason = goNogo(a)
-    println("")
-    @show decision
-    @show reason
-    # a.memory[:shortterm]["Check $(a.step):"] = reason
+    decision = "Yes"  # yes because a.step start at 0
+    if a.step != 0
+      decision, reason = goNogo(a)
+    end
+
     if decision == "Yes"   # in case there is a cancel, go straight to evaluation
       a.step += 1
-    elseif decision == "No"
-      # repeat the latest step
+    elseif decision == "No" # repeat the latest step
       a.memory[:shortterm] = removeHeaders(a.memory[:shortterm], a.step, ["Plan"])
       a.memory[:log] = removeHeaders(a.memory[:log], a.step, ["Plan"])
       println("repeating step $(a.step)")
     else
       error("undefined condition decision = $decision $(@__LINE__)")
     end
-    
-
     @show a.step
+    #WORKING checkStepCompletion
+    iscomplete = checkStepCompletion(a::agentReflex)
+    if iscomplete
+    
     if a.step < totalsteps   # the last step of the plan is responding, let work() do this part
 
       toolname, toolinput = actor_mistral_openorca(a)
@@ -933,7 +991,7 @@ function formulateUserresponse(a)
 
   prompt =
   """
-    <|im_start|>system
+    <|system|>
     Symbol:
     Plan: a plan
     Thought: your thought
@@ -948,8 +1006,9 @@ function formulateUserresponse(a)
     $work
     
     From your talk with the user and your work, formulate a response for the user.
-    <|im_end|>
-    response: 
+    <|/s|>
+    <|assistant|>
+    response:
   """
   response = sendReceivePrompt(a, prompt)
   return response
@@ -1080,7 +1139,7 @@ function goNogo(a)
 
   prompt =
   """
-    <|im_start|>system
+    <|system|>
     Symbol meaning:
     Plan: a plan
     Thought: your thought
@@ -1088,18 +1147,23 @@ function goNogo(a)
     Actinput: the input to the action
     Obs: the result of the action
 
-    Your work:
+    Your earlier work:
     $work
     
-    Your job is to check whether step $(a.step) of your work is completed according to the plan and choose only one of the following choices.
-    choice 1: If you get what you intend to do and you are ready to do the next step of the plan say, "{Yes}". And what is the rationale behind the decision to do the next step?
-    choice 2: If you didn't get what you intend to do and you need to repeat the latest step say, "{No}". And what is the rationale behind the decision to repeat the latest step?
-    <|im_end|>
-    <|im_start|>assistant
-
+    Your job is to check whether step $(a.step) of your work is completed according to the plan.
+    So for instance the following:
+    
+      step 2 of the plan: Ask user about the occasion type. But you can't find any relevant info of occasion type in your work.
+      assistant: Step 2 isn't done yet. {No}
+      step 5 of the plan: Ask user if they have any preference for the style of wine. And you found relevant info in your work such as the user like full-bodied wine.
+      assistant: Step 5 is done. {Yes}
+    </s|>
+    <|assistant|>
   """
 
   response = sendReceivePrompt(a, prompt)
+  @show goNogo_response = response
+
   decision = nothing
   reason = nothing
   if occursin("Yes", response)
@@ -1132,7 +1196,7 @@ function checkStepCompletion(a::agentReflex)
 
   prompt =
   """
-    <|im_start|>system
+    <|system|>
     Symbol meaning:
     Plan: a plan
     Thought: your thought
@@ -1144,8 +1208,8 @@ function checkStepCompletion(a::agentReflex)
     $plan
     
     What is step $(a.step) of the plan?
-    <|im_end|>
-    <|im_start|>assistant
+    <|/s|>
+    <|assistant|>
 
   """
 
@@ -1153,17 +1217,7 @@ function checkStepCompletion(a::agentReflex)
   
   response = split(response, "<|im_end|>")[1] 
 
-  # if occursin("N/A", response)
-  #   response = replace(response, "(N/A)"=>"")
-  # elseif occursin(response, "(not specified, assume casual)")
-  #   response = replace(response, "(not specified, assume casual)"=>"")
-  # else
-  # end
   
-  # mistral 7B already know info example: 2. Determine the occasion (wedding party).
-  if occursin("(", response) && occursin(")", response) 
-    result = true
-  end
   
   return result
 end
diff --git a/src/type.jl b/src/type.jl
index 643fb39..d42eeb8 100644
--- a/src/type.jl
+++ b/src/type.jl
@@ -168,7 +168,7 @@ function agentReflex(
       """
       Use the following format:
       Thought: you should always think about what to do according to step {step} of the plan (pay attention to correct numeral calculation and commonsense).
-      Act: the action to take that match your thought, should be one of [{toolnames}]
+      Act: the action to take that align with your thought, should be one of [{toolnames}]
       Actinput: your input to the action you chose (pay attention to the tool's input)
       Obs: the result of the action
       """,
diff --git a/src/utils.jl b/src/utils.jl
index a7a2f69..429a1fc 100644
--- a/src/utils.jl
+++ b/src/utils.jl
@@ -4,7 +4,7 @@ export makeSummary, sendReceivePrompt, chunktext, extractStepFromPlan, checkTota
         detectCharacters, findDetectedCharacter, extract_number, toolNameBeingCalled,
         isUseTools, conversationSummary, checkReasonableness, replaceHeaders,
         addShortMem!, splittext, dictToString, removeHeaders, keepOnlyKeys, experience,
-        messagesToString, messagesToString_nomark, removeTrailingCharacters
+        messagesToString, messagesToString_nomark, removeTrailingCharacters, dictLatestStep
 
 using UUIDs, Dates, DataStructures
 using CommUtils, GeneralUtils
@@ -378,35 +378,40 @@ Return:
 function isUseTools(a::agentReflex)
   toollines = ""
   for (toolname, v) in a.tools
-    if toolname ∉ ["chatbox"]
+    if toolname ∉ ["chatbox"] # LLM will always use chatbox
       toolline = "$toolname: $(v[:description]) $(v[:input]) $(v[:output])\n"
       toollines *= toolline
     end
   end
 
-  conversation = messagesToString_nomark(a.messages)
+  conversation = messagesToString(a.messages)
 
   prompt =
   """
-  <|im_start|>system
-  $(a.roles[a.role])
+  <|system|>
+  You are a helpful assistant.
   You have access to the following tools:
   $toollines
 
-  Your conversation with the user:
+  Your task is to decide whether you need think thoroughly in order to respond to the user according to your conversation with the user and tools you have. 
+  So for instance the following:
+  user: Hello!. How are you?
+  assistant: {No}, the user is greeting me, I could respond right away.
+  user: "I want a bottle of wine."
+  assistant: {Yes}, I need to think thoroughly about the user stimulus.
+  </s> 
   $conversation
-
-  From the conversation, ask yourself what do you intend to do now?
-  <|im_end|>
-  
+  <|assistant|>
   """
   
   # if LLM mentions any tools, use Plan/Thought/Act loop
   isusetool = false
-  response = sendReceivePrompt(a, prompt, temperature=0.0)
-  response = split(response, "<|im_end|>")[1]
+  response = sendReceivePrompt(a, prompt, temperature=0.2, max_tokens=64)
+  response = split(response, "<|assistant|>")[1] 
+  response = split(response, "<|user|>")[1]
+  @show response
   for (toolname, v) in a.tools
-    if occursin(toolname, String(response))
+    if occursin("Yes", String(response))
       isusetool = true
       break
     end
@@ -415,17 +420,14 @@ function isUseTools(a::agentReflex)
   if length(a.memory[:shortterm]) != 0
     isusetool = true
   end
-
   return isusetool
 end
 
 
-
-
 # function isUseTools(a::agentReflex)
 #   toollines = ""
 #   for (toolname, v) in a.tools
-#     if toolname ∉ ["chatbox"]
+#     if toolname ∉ ["chatbox"] # LLM will always use chatbox
 #       toolline = "$toolname: $(v[:description]) $(v[:input]) $(v[:output])\n"
 #       toollines *= toolline
 #     end
@@ -443,74 +445,30 @@ end
 #   Your conversation with the user:
 #   $conversation
 
-#   From your conversation, ask yourself what do you need to do now?
+#   From the conversation, ask yourself what do you intend to do now?
 #   <|im_end|>
   
 #   """
   
 #   # if LLM mentions any tools, use Plan/Thought/Act loop
 #   isusetool = false
-#   result = sendReceivePrompt(a, prompt, temperature=0.2)
+#   response = sendReceivePrompt(a, prompt, temperature=0.0)
+#   response = split(response, "<|im_end|>")[1]
 #   for (toolname, v) in a.tools
-#     if occursin(toolname, result)
+#     if occursin(toolname, String(response))
 #       isusetool = true
 #       break
 #     end
 #   end
-#   @show prompt
-#   whattodo = result
-#   @show whattodo
 
-#   return isusetool, whattodo
+#   if length(a.memory[:shortterm]) != 0
+#     isusetool = true
+#   end
+
+#   return isusetool
 # end
 
 
-# function isUseTools(a::agentReflex, usermsg::String)
-#     toollines = ""
-#     for (toolname, v) in a.tools
-#       if toolname ∉ ["chatbox"]
-#         toolline = "$toolname: $(v[:description]) $(v[:input]) $(v[:output])\n"
-#         toollines *= toolline
-#       end
-#     end
-
-#     prompt =
-#     """
-#     <|im_start|>system
-#     $(a.roles[a.role])
-#     You have access to the following tools:
-#     $toollines
-
-#     Your earlier conversation with the user:
-#     None
-
-#     User's message:
-#     $usermsg
-
-#     From the user's message, ask yourself what do you need to do?
-#     <|im_end|>
-#     <|im_start|>assistant
-
-#     """
-    
-#     # if LLM mentions any tools, use Plan/Thought/Act loop
-#     isusetool = false
-#     result = sendReceivePrompt(a, prompt, temperature=0.2)
-#     for (toolname, v) in a.tools
-#       if occursin(toolname, result)
-#         isusetool = true
-#         break
-#       end
-#     end
-
-#     whattodo = result
-#     @show whattodo
-
-#     return isusetool, whattodo
-# end
-
-
-
 
 """
   make a conversation summary.
@@ -614,9 +572,9 @@ function messagesToString(messages::AbstractVector{T}; addressAIas="assistant")
       end
 
       if role == "user"
-        conversation *= "<|im_start|>$role: $(content[1:end-nouse])\n<|im_end|>"
+        conversation *= "<|$role|>\n $(content[1:end-nouse])\n</s>"
       elseif role == "assistant"
-        conversation *= "<|im_start|>$addressAIas: $(content[1:end-nouse])\n<|im_end|>"
+        conversation *= "<|$addressAIas|>\n $(content[1:end-nouse])\n</s>"
       else
         error("undefied condition role = $role $(@__LINE__)")
       end
@@ -627,6 +585,36 @@ function messagesToString(messages::AbstractVector{T}; addressAIas="assistant")
 
   return conversation
 end
+# function messagesToString(messages::AbstractVector{T}; addressAIas="assistant") where {T<:AbstractDict}
+#   conversation = ""
+#   if length(messages)!= 0
+#     for msg in messages
+#       role = msg[:role]
+#       content = msg[:content]
+      
+#       nouse = 0
+#       for i in reverse(content)
+#         if i == '\n' || i == ' '
+#           nouse += 1
+#         else
+#           break
+#         end
+#       end
+
+#       if role == "user"
+#         conversation *= "<|im_start|>$role: $(content[1:end-nouse])\n<|im_end|>"
+#       elseif role == "assistant"
+#         conversation *= "<|im_start|>$addressAIas: $(content[1:end-nouse])\n<|im_end|>"
+#       else
+#         error("undefied condition role = $role $(@__LINE__)")
+#       end
+#     end
+#   else
+#     conversation = "N/A"
+#   end
+
+#   return conversation
+# end
 
 """ Convert a vector of dict into 1-continous string.
 
@@ -1028,6 +1016,14 @@ function experience(dict::T) where {T<:AbstractDict}
 end
 
 
+function dictLatestStep(dict::T) where {T<:AbstractDict}
+  @show dict
+  _latest_step = keys(dict)
+  _latest_step = [i for i in _latest_step]
+  _latest_step = _latest_step[end]
+  latest_step = parse(Int, _latest_step[end-2:end-1])
+  return latest_step
+end