While making social media creatives for clients. AI tools design them fine now, but the output is one flat image. So when there's a typo or a wrong price (there's always something), you can't fix it. You regenerate and hope the good parts survive. I got sick of this.
So I made this. Please try it- [GITHUB](https://github.com/Abirhossainzozo/layerly-creatives)
It's a skill for Claude / ChatGPT / Codex / Cursor / Gemini. The model doesn't output an image, it outputs a layout file. A Python script renders each element to its own layer and checks for overlapping text, bad contrast, stuff going off canvas. Then a Node script (vendored ag-psd) writes an actual PSD with editable text layers, masks, shadows, folders. Opens in Photoshop or Photopea.
The messy part was rebuilding existing flat images as layers. You can't un-merge a JPG. So it keeps people as one layer, removes loose objects with generative fill, regenerates those as cutouts, and has to show me a sheet of every cutout on grey before it's allowed to deliver. Added that after it kept lying to me about "clean" cutouts.
There are before/after examples in the repo from my actual client work, PSDs included, so you can open them and check the layers yourself.
About the license since HN cares: free for personal use (PolyForm Noncommercial), paid for business use. So source-available, not open source.
I'm not an expert. I just kept hitting this problem. Tell me what breaks, criticism welcome.