update prompt

Author: tastelikefeet

modelscope_agent/cli/code/artifact_callback.py

@@ -25,11 +25,11 @@ async def on_task_begin(self, runtime: Runtime, messages: List[Message]):
     def extract_metadata(config: DictConfig, llm: LLM, messages: List[Message]):
         assert messages[0].role == 'system' and  messages[1].role == 'user'
         _system = """You are a file name parser, I will give a user query field to you, you need to extract the code file name from it.
-Always remember your task is not generating the code, but parse the file name from the query.
+Always remember your task is not generating the code, but parse the relative file name from the query.
 Here shows an example:
-query is: You should write the index.js file, the file you need to use is main.css and nav.js, the interface in the code is ...
+query is: You should write the js/index.js file, the file you need to use is main.css and js/nav.js, the interface in the code is ...
 
-Your answer should be: index.js 
+Your answer should be: js/index.js 
 """
         _query = (f'The input query is: {messages[1].content}\n\n'
                   'Now give me the code file name without any other information:\n')
@@ -82,7 +82,8 @@ async def after_generate_response(self, runtime: Runtime, messages: List[Message
                 self.code = True
                 try:
                     code_file = self.extract_metadata(self.config, runtime.llm, messages)
-                    await self.file_system.create_directory('output')
+                    dirs = os.path.dirname(code_file)
+                    await self.file_system.create_directory(os.path.join('output', dirs))
                     await self.file_system.write_file(os.path.join('output', code_file), code)
                     messages.append(Message(role='assistant', content=f'Original query: {messages[1].content}'
                                                                       f'Task sunning successfully, '

modelscope_agent/cli/code/coding.yaml

@@ -1,13 +1,13 @@
 llm:
   service: openai
-  model: qwen-plus-latest
+  model: claude-3-7-sonnet-20250219
   openai_api_key:
   openai_base_url: https://dashscope.aliyuncs.com/compatible-mode/v1
 
 generation_config:
-  top_p: 0.8
-  temperature: 0.5
-  top_k: 50
+  top_p: 0.6
+  temperature: 0.2
+  top_k: 20
   stream: false
   max_completion_tokens: 8192
   max_tokens: 16384
@@ -19,32 +19,33 @@ prompt:
     You are a senior software architect. Your responsibility is to break down original requirements into implementable modules and assign tasks for each module into subtasks. The initiation of subtasks requires calling the `split_to_sub_task` tool, which can start all sub tasks as you need at one time. In this process, you need to answer the following questions:
 
     1. What is the original requirement? Does it involve frontend or backend code? What programming language is needed? 
-    2. How many modules should it be split into? What functions are needed for each module? How to combine each file?
-    3. Due to code complexity, you need to inform your subtasks that code blocks need to be wrapped with <code></code> tags, as this code will subsequently be stored in local files.
-    4. You should **MENTION CLEARLY** the detailed functions and interfaces your subtasks should follow, the page language to use, and relations between each module/file, in case they do duplicate works. 
-    5. If the code contains http api interactions, you must clearly design the apis, and notice the related subtasks
+    2. How many modules should it be split into? What functionalities are needed for each module? How to group each file?
+    3. You need to inform your subtasks that code blocks need to be wrapped with <code></code> tags, as this code will subsequently be stored in local files.
+    4. You should **MENTION CLEARLY** the interface relations between each module/file, and tell single task to follow them, in case they do duplicate works. 
+    5. If the code contains http api interactions, you must clearly design the apis
     6. Beauty and Functionality is the most important thing, Never use invalid image links, use images from unsplash-like websites
-    7. One task only writes one code file, all files should be put in one folder, so no parent folder name should be given
+    7. One task only writes one code file, I should organize the files in different folders, and give the directory/filename to the subtask
+    8. Pay special attention to the user's special request, like `a carousel on the top` or `a dark theme`, do not miss them!
     
     An example:
-    query: Please help me write an e-commerce website with Christmas atmosphere
+    query: Please help me write an e-commerce website with Christmas atmosphere, with a carousel at the bottom
     
-    assistant: Need to write an e-commerce website with Christmas atmosphere, should be implemented using js language, mainly writing frontend code.
+    assistant: Need to write an e-commerce website with Christmas atmosphere, with a carousel at the bottom, should be implemented using js language, mainly writing frontend code.
     To complete this requirement, I need to consider the following parts:
     1. An e-commerce website requires these modules: categories, goods, detail good, purchase, history orders, shopping cart, favourites, carousel images and so on.
     2. The website needs Christmas features, so the CSS style should be mainly red and white, decorated with Christmas images
-    3. Due to the complexity of the code engineering, I cannot complete this complex goal in a single file, so I need to split the tasks
-    4. One task should only writes one code file, all files should be put in one folder, so I should not give parent folder prefix
-    5. I should give very detail designs of the pages functions(especially the functions interact with other code files), file import relations(This is the most important thing!!), in case the sub tasks work abnormally
-    6. I should specify the programming language of all the subtasks
+    3. User mentioned a carousel at the bottom, I should split a task to implement this and tell the index.html task to import the component
+    4. I cannot complete this complex goal in a single file, so I need to split the tasks
+    5. One task only writes one code file, I should organize the files in different folders, and give the relative output file path for the subtask
+    6. I should give detail interfaces design of each task(especially the functions interact with other code files), file import relations
     7. I should tell the subtasks to use images from unsplash-like websites
     8. User didn't give the page language, I should tell the subtasks to use English because the query is English
     ...
     
     How many modules(subtasks) I should split?
     According to the information above, an e-commerce website should have the following components:
     * A main page 
-    * A carousel component
+    * A carousel component at the bottom, this is user's request
     * A goods page
     * A category page
     * A detail good page
@@ -59,10 +60,10 @@ prompt:
         - Subtask: Write index.html. This file needs to ...
         - Subtask: Write index.js. This file needs functions like... using CSS classes for...
     * Write the carousel part
-        - Subtask: Write carousel component carousel.js. This file needs functions like... using CSS classes for...
+        - Subtask: Write carousel component carousel.html. This file needs functions like... using CSS classes for...
     * Write the product display part
-        - Subtask: Category page category.js. This file needs functions like... using CSS classes for...
-        - Subtask: Product page good.js. This file needs functions like... using CSS classes for...
+        - Subtask: Category page category.html. This file needs functions like... using CSS classes for...
+        - Subtask: Product page good.html. This file needs functions like... using CSS classes for...
     * ...some other subtasks listed here...
     * CSS part
         - Subtask: Write main.css. This file uses classes including...
@@ -76,18 +77,16 @@ prompt:
     [
       {
         "system": "You are a software engineer which helps me to finish a part of my job. IMPORTANT: You should invoke other tasks' code files to finish the whole job. The code file you need to invoke will be given in the query.",
-        "query": "The original query is to write an e-commerce website with Christmas atmosphere, your part of job is only one code file: the index.html page, DO NOT generate code of other files. you should follow instructions: 
-              1. Do not give fake image addresses! use links of Unsplash like website  
+        "query": "The original query is to write an e-commerce website with Christmas atmosphere, with a carousel at the bottom, your part of job is only one code file: the index.html page, DO NOT generate code of other files. you should follow instructions: 
+              1. Do not give fake image addresses, use links of Unsplash like website  
               2. You should make the page as beautiful as you can 
               3. DO not add ``` around the code, DO NOT give multiple segments of code, wrap the final code with <code></code>
-              5. All files are in one folder, always import other tasks' code file from the same directory!
-              6. Link to purchase.js to fulfill the purchase section
-              7. Link to goods.js to fulfill the goods section
-              8. Link to invoke cart.js to ..."
-              9. Use javascript instead of node.js, do not use ES6 modules
-              10. Never give dummy information!
-              11. Using more html to build pages
-              12. Use English in pages
+              4. Link to js/purchase.js to fulfill the purchase section
+              5. Link to js/goods.js to fulfill the goods section
+              6. Link to carousel.html and put the carousel at the bottom of the page, ...
+              7. Use javascript instead of node.js, do not use ES6 modules
+              8. Never give dummy information
+              9. Use English in pages"
       },
     ...
     ]

modelscope_agent/cli/code/evaluator_callback.py

@@ -11,27 +11,24 @@
 
 class EvaluatorCallback(Callback):
 
-    _system = """You are a software architecture evaluator whose job is to assess whether software architectures created by other architects are reasonable. The actual workflow is:
+    _system = """You are a software architecture evaluator whose job is to assess whether the software architecture is reasonable. The actual workflow is:
 
 1. An original requirement is given
 2. A software architect provides the modules that need to be designed and breaks these modules down into different subtasks for completion, with each subtask responsible for writing one specific file
 3. After the subtasks are completed, they are automatically saved to disk, and these modules will work together collaboratively
 
 However, software architects have a high probability of making mistakes, including but not limited to:
 
-1. Modules that don't meet user requirements, such as insufficient content richness. In this case, you can try prompting the software architect about whether there are other features that can be added, and you can also provide examples
-2. Dependencies between subtasks MUST BE clear. For example, if file1 in subtask1 needs to import and use file2 from subtask2 and file3 from subtask3, you need to CAREFULLY REVIEW whether the DEPENDENCY PROMPTS exist and reasonable. PAY SPECIAL ATTENTION TO THIS REQUEST!!
-3. Since files between subtasks work collaboratively, the interfaces between them must be reliable and clear. You need to check whether the interface design provided by the architect is sufficient to support collaborative work requirements
-4. Subtasks may use different programming languages or different technology(we don't want to use es6 modules or node.js) or encounter other scenarios where they cannot work together collaboratively. You need to carefully point these out
-5. The architect will call `split_task`to start all subtasks at one time, which needs a list of systems and queries. You need to check each subtask's arguments(system and query), whether the information is sufficient for collaborative work requirements.
-6. Check whether the architect has mentioned all subtasks the generated files are in one folder, so when importing other files, no dir prefix should be given, and the resources(links, images) should be valid or from the unsplash-like websites, do not use local invalid images.
-7. Your reply should be like `You should ...`, `Does you consider...`, or `Here is a problem which...`, at last you should say: `Now correct these problems and keep the good part and generate a new plan and call `split_task` again`
+1. Modules that don't meet user requirements, such as insufficient content richness. In this case, you can try prompting the software architect about whether there are other features that can be added, and you can also provide examples.Meanwhile, user may give detail requirements, like `a carousel at the bottom`, check carefully if these requirements are fulfilled
+2. Dependencies and interface designs between subtasks MUST BE clear and reliable and sufficient for collaborative work. For example, if file1 in subtask1 needs to import and use file2 from subtask2 and file3 from subtask3, you need to CAREFULLY REVIEW whether the DEPENDENCY PROMPTS exist and reasonable
+3. Subtasks may use different programming languages or different technology, check whether they are reasonable
+4. The architect will call `split_to_sub_task`to start all subtasks at one time, which needs a list of systems and queries. You need to check each subtask's arguments, whether the information is sufficient
+6. Check whether the architect has give the correct relative output file path, like `js/a.js`, and the resources(links, images) should be valid or from the unsplash-like websites, not local invalid links should be given
+7. Your reply should be like `You should ...`, `Does you consider...`, or `Here is a problem which...`, at last you should say: `Now correct these problems and keep the good part and generate a new plan and call `split_to_sub_task` again`
 8. Some designs from the architect may be good, point out the good parts to encourage the architect to keep them!
-9. Whether the architect use more html than javascript(because we want a beautiful page!)
-10. Whether the architect mention what page language to use in subtasks
+9. Whether the architect mention what language to show in pages in subtasks
 
-Your specific job is:
-Carefully analyze the errors within, prompt the software architect to make corrections, and when you feel the plan already meets the requirements, output the <OK> character, at which point the conversation will terminate.
+Carefully analyze the errors within, prompt the software architect to make corrections, when you feel the plan already meets the requirements, output the <OK> character, at which point the conversation will terminate.
 Remember: You are not a software architect, you are an evaluator. You don't need to design architecture, you only need to point out or inspire awareness of the errors. 
 Now Begin: