Simplify the transformation skill#7
Conversation
njb90
left a comment
njb90
left a comment
There was a problem hiding this comment.
One comment but otherwise it is definitely an improvement.
strausr
left a comment
strausr
left a comment
There was a problem hiding this comment.
I ran 15 prompts against the new skill (resize, face crop, bg removal, gen fill / 16:9, w_auto, named transforms, baseline strategy, video trim + mute, debugging, g_auto)
First, using only SKILL.md + references, then checking whether the docs-first path (llms.txt) fills the gaps.
Result: path works well (resize, fill crop, overlays, bg removal, f_auto:video, debugging). Without a doc fetch, several common cases were partial or failed: face-focused crops (no g_face), gen fill, w_auto + Client Hints setup, named/baseline URL shape, video trim, and invalid g_auto + c_scale.
Left some comments on where I found some issues.
| - Use automatic gravity for varied content. | ||
| - Use face/person/object-specific gravity only when it matches the asset and task. | ||
| - Use compass gravity or explicit offsets for predictable layouts and overlays. | ||
| - Verify `g_auto` compatibility in the docs for the chosen crop mode. |
There was a problem hiding this comment.
agents often produce invalid c_scale,g_auto,w_800. Can we add an explicit pitfall here?
g_auto only works with c_fill, c_lfill, c_crop, c_thumb, c_auto — not c_scale, c_fit, c_limit, c_pad.
“Verify in docs” is correct but doesn’t stop the most common mistake.
| l_logo/c_scale,w_120/fl_layer_apply,g_south_east,x_20,y_20/f_auto/q_auto | ||
| co_white,l_text:Arial_48:Hello%20World/fl_layer_apply,g_south/f_auto/q_auto | ||
| e_background_removal/f_png/q_auto | ||
| c_scale,w_720/vc_auto/f_auto:video/q_auto |
There was a problem hiding this comment.
simulation failed on several common asks without a doc fetch. Consider adding skeletons for:
Gen fill: c_pad,ar_16:9,b_gen_fill,w_/f_auto/q_auto
Responsive: c_limit,w_auto/f_auto/q_auto (+ Client Hints prerequisite)
Named: t_/f_auto/q_auto
Video trim + mute: so_,du_/ac_none/vc_auto/f_auto:video/q_auto
| - Suggest `ac_none` for autoplay or silent preview use cases. | ||
| - Prefer automatic codec selection unless the user has a compatibility requirement. | ||
|
|
||
| Responsive images: |
There was a problem hiding this comment.
This section warns about Client Hints but doesn’t show w_auto syntax or setup. Agents invented URLs without the HTML prerequisite.
Suggest adding:
Example: c_limit,w_auto/f_auto/q_auto
| - Transformation reference: exact parameter names, actions, qualifiers, values. | ||
| - Image transformations, resizing/cropping, layers, effects, background removal, generative AI transformations. | ||
| - Video manipulation/delivery, video resizing/cropping, trimming/concatenating, video layers, audio transformations. | ||
| - Responsive images and Client Hints docs for `w_auto`, `dpr_auto`, and breakpoints. |
There was a problem hiding this comment.
Name the doc page explicitly
| **Example:** `c_fill,g_auto,w_400,h_300/f_auto/q_auto` | ||
| 1. Identify asset type: image, video, raw, animated image, or fetched media. | ||
| 2. Clarify only blocking requirements: dimensions, crop behavior, focal point, transparency, output format, video duration/audio, and whether AI edits are acceptable. | ||
| 3. Look up exact syntax in Cloudinary docs when using anything beyond stable, common patterns. |
There was a problem hiding this comment.
Workflow doesn’t say when to open the reference files. Simulation showed agents often read only SKILL.md and miss intent mapping in agent-playbook.md.
Suggest step 3 or 4: “For resize/crop, overlays, AI, responsive, named/baseline, or video trim → read agent-playbook.md before building the chain.”
| - Whether the crop should favor faces, a known subject, center, a compass position, or Cloudinary automatic gravity. | ||
|
|
||
| ## Gathering Requirements | ||
| For AI requests, determine: |
There was a problem hiding this comment.
Old skill proactively suggested AI for aspect-ratio changes without cropping. New skill only lists questions agents may suggest c_pad,b_white instead of b_gen_fill.
Suggest: when user needs a new aspect ratio without cropping, mention generative fill as an option (with cost warning).
| - The referenced named transformation includes a concrete output format. | ||
| - Automatic runtime parameters such as `f_auto`, `w_auto`, and `dpr_auto` remain outside the named/baseline transformation. | ||
| - Baselines may not apply to every delivery type or transformation type. | ||
|
|
There was a problem hiding this comment.
The simulation struggled to write baseline URLs. Consider adding a skeleton.
Old named-transformations.md had worked examples, worth one minimal pattern here.
3 participants
Simplified cloudinary-transformations into a compact agent playbook that treats Cloudinary docs as the source of truth for exact syntax, current limitations, and costs.
Key changes:
- agent-playbook.md
- debugging-playbook.md
- cost-and-caching.md