Installs into .claude/skills of the current project.
Are you the author of Youtube Profile Spec?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/lorena-bordonaba-pau-youtube-profile-spec)
---
name: youtube-profile-spec
description: Specification for a YouTube profile picture — 1:1, cropped to a circle. Triggered when generating or reviewing a channel profile picture.
---
# Profile picture specification
## WHEN
Any profile picture generation or review.
## THE RULE THAT DECIDES EVERYTHING
It is uploaded **square, 1:1**, but YouTube **crops it to a circle**. Anything
near the corners is lost.
And it is seen small: **32x32 pixels in the comments**. That is the real size
it has to be recognisable at, not the one in your editor.
## EXECUTION ORDER
1. **See the current one**
`python3 tools/view_channel_packaging.py --only avatar --md` and open it.
2. **Generate**
```
python3 tools/generate_image.py --type profile_image --prompt "..." \
--ref face_photo.jpg:likeness --dry-run
```
`--type profile_image` already asks for a centred subject, a simple
background, no text and legibility at 32x32.
If the creator's face appears, load `likeness-preservation`: in an avatar
the identity is all there is.
3. **Verify at real size**, which is what almost nobody does:
`python3 tools/export_image.py --image PATH --size 32x32 --out /tmp/avatar32.png`
and open that 32x32 file. If you cannot tell who it is at that size, it does
not work, however good it looks at 800x800.
4. Export the final one:
`python3 tools/export_image.py --image PATH --type profile_image`
## COMPOSITION
- Subject centred and large: the circle eats the corners.
- Simple, single-colour background, high contrast against the subject.
- **No text and no fine detail.** At 32x32 it is a smudge.
- Colour-consistent with the banner.
## SOURCES
`source: generated` for the image; `derived` for the resizing.
## OUTPUT
1. The final profile picture, exported with `--type profile_image`.
2. The 32x32 check: whether it is still recognisable at comment size.
3. The model that generated it, named explicitly.