Generating scripts with LLMs

With the recent release of the new OpenAI ChatGPT that can do image generation, I thought that I might add a feature image to my last blog post. I’ve had mixed experiences of AI generation of images, probably because I’m not good at prompting them.

I started with this:

We now have an image, but it doesn’t have a credit. Now, I could go back to the LLM and ask it again, but these things are slow and I won’t get the same image anyway. So I decided to write a python script to add an image credit to the bottom right of the image.

Clearly this is a job for an LLM, so lets use Claude Code.

I started it up and gave it this prompt:

> Create a single file python script called add-image-credit that has executable permissions. This file starts with the line 
  `#!/usr/bin/env -S uv run 
    --script --quiet` and uses the `# /// script` system to bring in any dependencies.

  This script will take an image filename and a credit text string. Display the credit text in white with a drop shadow in 
  the bottom right hand corner of the image and save the image with the same name with the word "credited" appended. The text
   font size should be no more than 18, but small enough such that the string is no more than 30% of the width of the image. 
  Use Helvetica as the font, but make this a variable so that it can be easily changed. If the credit text is not supplied, 
  then use the text "Image generated by AI" by default. 

Claude does it stuff:

⏺ I'll create a Python script that adds credit text to images as specified. Let me write the script with executable
  permissions.

⏺ Write(file_path: add-image-credit)…
  ⎿ Wrote 80 lines to add-image-credit
     #!/usr/bin/env -S uv run --script --quiet
     # /// script
     # /// dep pillow
     import argparse
     import os
     from PIL import Image, ImageDraw, ImageFont
     def add_credit(image_path, credit_text=None, font_name="Helvetica"):
         if credit_text is None:
     … +70 lines (ctrl+r to see all)

⏺ Bash(chmod +x /Users/rob/Projects/image-credit/add-image-credit)…
  ⎿  (No content)

⏺ The script has been created with executable permissions. It:
  - Uses PIL to add credit text to images
  - Automatically scales font size (max 18pt) to ensure text width is ≤30% of image width
  - Places white text with black drop shadow in bottom right corner
  - Saves output as original_filename_credited.ext

  Run it with: ./add-image-credit image.jpg --credit "Your credit text"

Interestingly, it’s got the format for Inline script metadata incorrect, but I didn’t notice and asked it to run it.

> Run this script against backing-up-signal-feature-image.png that is in this directory

⏺ I'll run the script against the PNG file. Let me first check if the file exists and then run the script.

⏺ List(path: "")…
  ⎿  - /Users/rob/Projects/image-credit/
       - add-image-credit
       - backing-up-signal-feature-image.png
     … +1 item (ctrl+r to see all)

⏺ Bash(./add-image-credit /Users/rob/Projects/image-credit/backing-up-signal-feature-image.png)…
  ⎿  error: An opening tag (`# /// script`) was found without a closing tag (`# ///`). Ensure that every line between the 
     opening and closing tags (including empty lines) starts with a leading `#`.

The script has failed due to the error related to the Inline script metadata, Claude detects and fixes it:

⏺ I need to fix the script format. The script tag is missing the closing marker.

⏺ Write(file_path: add-image-credit)…
  ⎿  Updated add-image-credit with 1 addition
     1  #!/usr/bin/env -S uv run --script --quiet
     2  # /// script
     3  # /// dep pillow
     4  # ///
     5 
     6  import argparse
     7  import os

⏺ Bash(./add-image-credit /Users/rob/Projects/image-credit/backing-up-signal-feature-image.png)…
  ⎿  error: TOML parse error at line 1, column 1
       |
     1 | /// dep pillow
       | ^
     invalid key

⏺ Let me fix the script format again. There seems to be an issue with the uv script dependency format.

⏺ Write(file_path: add-image-credit)…
  ⎿  Updated add-image-credit with 1 addition and 1 removal
     1  #!/usr/bin/env -S uv run --script --quiet
     2  # /// script
     3  # /// dep pillow
     3  # dependencies = ["pillow"]
     4  # ///
     5 
     6  import argparse

Now it has got it right, but it also got something wrong with colours, so had to fix them too. I’ve skipped that iteration here as it’s the same loop as we’ve just done.

⏺ Bash(./add-image-credit /Users/rob/Projects/image-credit/backing-up-signal-feature-image.png)…
  ⎿  Credited image saved as: /Users/rob/Projects/image-credit/backing-up-signal-feature-image_credited.png

⏺ Great! The script now successfully adds the default credit text "Image generated by AI" to the image.

It was successful and I quit Claude:

> exit



Total cost:            $0.2396
Total duration (API):  1m 40.8s
Total duration (wall): 5m 36.3s
Total code changes:    99 lines added, 7 lines removed

Not bad for 24 cents and I now have a new script that I’ve moved to ~/bin!

This is the final image:

For one-off things like this, it’s more than good enough. Letting it loose on a bigger codebase will cost a lot more as the amount of code sent to Claude and processed by it will significantly affect the price. Also, any code created needs testing quite thoroughly and if you’re going to put it into production, it need to be looked at and assessed carefully.