Edit an image that already exists instead of regenerating it from scratch. Triggered by "change the background", "remove the logo", "make it darker", "same but without the text". In Spanish, "cámbiale el fondo", "quítale el logo", "hazlo más oscuro", "igual pero sin el texto".
Installs into .claude/skills of the current project.
Are you the author of Image Refinement?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/lorena-bordonaba-pau-image-refinement)
---
name: image-refinement
description: Edit an image that already exists instead of regenerating it from scratch. Triggered by "change the background", "remove the logo", "make it darker", "same but without the text". In Spanish, "cámbiale el fondo", "quítale el logo", "hazlo más oscuro", "igual pero sin el texto".
---
# Image refinement
## WHEN
There is an image that is already mostly right and one part needs changing.
**The mistake this skill exists to prevent**: passing the nearly-good image as
a *reference* to `generate_image`. The model then draws another similar image
and whatever was working is lost. The image to edit goes in as the **source**,
never as a reference.
## BEFORE SPENDING A CREDIT
If the change is about **size, crop, format or weight**, this is not an edit:
`python3 tools/export_image.py --image PATH --type thumbnail`
It is local, deterministic and free. It redraws nothing.
## EXECUTION ORDER
1. **Open the current image** with the image reading tool and name exactly what
is wrong. Without that, the instruction comes out vague and so does the
result.
2. **One instruction, imperative, only the change**
```
python3 tools/refine_image.py --image PATH \
--instruction "change the background to flat cobalt blue"
```
The tool already adds *"keep everything else exactly as it is"*. Describing
the whole image in the instruction is what makes it redraw.
`--type` defaults to `preserve`, which keeps the original's shape. Only
change it when a different shape was actually asked for.
3. **If there is a face**, pass it as a separate reference:
`--ref face_photo.jpg:likeness`. The face does **not** go in
`--instruction`.
4. **Compare source and result by opening both.** An editing model changes
things it was not asked to more often than you would expect.
5. Iterate one change at a time. Two changes in the same instruction collide.
## SOURCES
`source: generated`. An artefact, not a measurement. Name the model when
delivering.
## OUTPUT
The file path, the model that made it, and **what changed compared to the
source, after looking at both** — not an assumption about what should have
changed.
