Code Assistant Tests

The Task

Build small web app that accepts an image as input through an UI and returns a description of the image. We'll tackle this task with:

• Continue (locally hosted models),
• Github Copilot,
• Windsurf,
• and Cline (locally hosted models).

We'll use Brief Product Requirements (click to see) as custom instructions for the code assistant within .continuerules, .github/copilot-instructions.md, .windsurfrules and .clinerules.

import ollama

response = ollama.chat(
    model='llama3.2-vision',
    messages=[{
        'role': 'user',
        'content': 'What is in this image?',
        'images': ['image.jpg']
    }]
)

print(response)

root/
  backend/
    main.py
    pyproject.toml
    ...
  frontend/
    package.json
    ...

Prompts I used

Example prompt (so that you can read the conversation history better):

Write a haiku

I do the setup myself to not distract the model and to set some structure. Note: for each ai tool reference to files/folders via '@' or '#'. Don't just copy the prompt. Should help the model understand the context better.

So for a replay, clone the repo and cd into the ai-tool folder and open VSC (Continue, Copilot, Cline) or Windsurf.

Note, for some tools follow-up commands & prompts were necessary. I documented them in corresponding README.md files.

Preparations

# Install llama3.2-vision (if you don't already have it)
ollama pull llama3.2-vision

# Move into the ai-tool folder
cd <INSERT_TOOL_NAME> # e.g. cd copilot

# Create backend
mkdir backend && cd backend
uv init && uv add fastapi ollama
mv hello.py main.py

# Move test_image.png into the ai-tool/backend folder
cp ../../test_image.png .

# Create frontend (Yes to all defaults)
cd ..
npx create-next-app@latest frontend

1. File Upload

Initial prompt:

I created the nextjs project within frontend/. Help me implementing the File Upload feature. Note, we'll work on backend/ after that.

2. Image to Text

Initial prompt:

We've created a nice looking UI. Now let's move on to Image to Text. I initialized the uv project first and installed FastAPI for the backend (cd backend && uv init && uv add fastapi). Create the fastapi backend. It should accept an image as input from the frontend, send the image to the ollama model (you have an example for this) and receive the response from ollama. I also provided a test_image.png in backend/.

3. Result Display

Frontend seems to work and the backend as well (both are turned off at the moment). Now let's work on Result Display.

Follow up prompt (bonus):

In general it works. But it would also be nice to see the image which was uploaded. Let's add that.

Results & Conclusion

Every tool had the same initial data and prompts. However, some tools seemed to understand better what I wanted. I was a bit surprised, that Github Copilot Edits in combination with o1 performed so much worse than Claude in Windsurf. I assume it's not just the model, but the context and how it is processed. It could also be that my .*rules had too many characters.

Windsurf ✅

• Holy!
• Fast iterations. Fixed issues automagically.
• Created a nice looking web app that did what I wanted.

Github Copilot ✅

• Took a bit longer. But it worked.
• Together we could fix some issues. Didn't get stuck.
• NOTE: I used Copilot Edits which is currently still a preview version.

Continue ❌

• Macbook got loud sometimes. Had to brew services restart ollama or stop ollama completely (100% GPU).
• In the chat ENTER was my friend. CMD+ENTER confused the model since it checked the whole codebase.
• The backend was created without any issues. Frontend didn't work as expected. Couldn't upload a file. Not sure what the issue was. Don't have much experience with it. Gave up.

Cline ❌

• Same as Continue. Mac gets loud quickly with too much context
• No autocomplete like Copilot or Continue. Maybe combo Cline & Continue does the trick?
• Didn't even get the first prompt to work. Divide & Conquer wasn't possible. Cline always wanted to create frontend & backend at once.
• Context in .clinerules was always repeated. Yet, deleting and adding details into the prompt didn't help either.
• Gave up.