Automatizacija: AI vizualas skirtingiems formatams (ChatGPT + Canva pamoka)

@Canva prisijunk prie mano „Canva“ paskyros.

@Canva parodyk mano naujausius „Canva“ dizainus.

Create and install a personal Skill named `canva-new-generate`.

Create these exact files:

1. `SKILL.md`
2. `references/qa.md`

Use the following content without shortening or replacing the workflow with Canva's own AI design generator.

--- SKILL.md ---

---
name: canva-new-generate
description: Regenerate existing Canva designs or source images as native compositions for new aspect ratios, then convert each result into an editable Canva design with Magic Layers, repair OCR and Lithuanian glyphs, and visually verify the actual rendered output. Use when the user invokes @canva-new-generate, asks for fresh 1:1, 16:9, 9:16, 4:5, or custom-ratio Canva versions based on an original visual, wants regeneration instead of Canva resize, or needs a major portrait-to-landscape or landscape-to-portrait adaptation.
---

# Canva New Generate

Create a fresh image for every requested ratio using the original as a visual reference, then make it editable with Canva Magic Layers. Optimize for native composition and visual quality rather than preserving the original layer coordinates.

## Non-negotiable behavior

- Preserve the original source and existing Canva designs.
- Do not use Canva `resize-design`, `prepare-design-generation`, or `generate-design`.
- Generate one independent image per target ratio; never crop or stretch one generated result into the other formats.
- Use the original only as a reference for message, character, palette, objects, typography mood, and visual identity.
- Inspect the actual source image, generated image, and post–Magic Layers render. Never claim visual QA from layer coordinates or API success alone.
- Keep every requested result editable through Magic Layers, while clearly reporting any grouped raster elements.
- Proceed without outline or intermediate-save questions. Pause only for a missing source or Canva's currently required final save approval.

## 1. Resolve and inspect the source

1. Prefer the user's attached or local source image.
2. For a Canva source, resolve the design, retrieve a current full-page render, and obtain the highest-quality usable reference available.
3. Load the reference with the local image viewer before generation.
4. Record:
   - exact text and intentional line breaks;
   - primary and supporting hierarchy;
   - recurring character and object traits;
   - palette, lighting, materials, and mood;
   - elements that must remain separate;
   - flaws that must not be reproduced.
5. Treat text, dates, names, numbers, and Lithuanian diacritics as locked invariants.

## 2. Plan each format natively

Use standard dimensions when the requested format is unambiguous:

| Ratio | Default canvas |
| --- | --- |
| 1:1 | 1080 × 1080 |
| 4:5 | 1080 × 1350 |
| 16:9 | 1920 × 1080 |
| 9:16 | 1080 × 1920 |

Honor an explicitly requested unusual ratio exactly.

Design for the new orientation:

- **Landscape:** use horizontal zones or columns and intentionally ground people or products.
- **Portrait:** use a vertical narrative with clear stacked sections.
- **Square:** rebalance the hierarchy to use the full width and height.

Design for Magic Layers: separate foreground objects, use simple text areas, avoid text across detailed imagery, limit unnecessary overlap, and keep icons clear.

## 3. Generate from the reference

1. Use the image-generation tool with the source as a reference image.
2. Prompt for a fresh native composition at the exact target ratio, not an edit that merely expands or crops the source.
3. Include the exact on-image text verbatim and require complete Lithuanian Latin Extended support when applicable.
4. State the required relationships explicitly:
   - which side contains the headline;
   - where the subject is grounded;
   - which icons flank a text card;
   - which elements must never overlap;
   - safe margins and reading order.
5. Generate formats separately, one call per ratio.

## 4. Inspect the generated image

Open the actual generated workspace image at original detail. Check:

- exact ratio and native-looking composition;
- spelling, punctuation, line breaks, and `ą č ę ė į š ų ū ž` plus uppercase forms;
- hands, faces, limbs, object count, and perspective;
- intentional subject crop and grounding;
- headline, icon, card, and CTA separation;
- no empty dead zone, collision, clipped text, logo, watermark, or invented copy.

Regenerate once with a targeted correction when a failure is material. Do not send a visibly flawed source into Magic Layers.

## 5. Convert with Magic Layers

For every approved generated image:

1. Call Canva `image-to-design` separately with the final workspace image path.
2. Use a clear title containing the subject and ratio.
3. Start an editing transaction and audit every extracted text and image layer.
4. Compare each editable text layer with the locked source copy.
5. Correct all OCR substitutions, capitalization, punctuation, and line breaks. Check explicitly for:
   - `ė` → `é`;
   - `ž` → `ż`;
   - `į` → `j` or `i`;
   - `š` → `s`, `ş`, or dotless forms;
   - `ū` → `u`, `ú`, or `ü`.
6. Recalculate positions after any text-box or font-size change.
7. Record how many editable text layers, image layers, and grouped raster elements remain.

## 6. Run real rendered QA

Read [references/qa.md](references/qa.md) and apply every relevant check.

After every correction batch:

1. Request a fresh full-page Canva preview.
2. Show the preview as required by the connector.
3. Download that exact returned preview to the workspace.
4. Open it with the local image viewer.
5. Compare the rendered image with the intended composition and returned geometry.
6. Fix every visible collision, crop, imbalance, text-fit, or glyph problem and render again.

Never substitute a coordinate audit for visual inspection. Stop after three materially different correction passes if the design cannot converge and report the specific Magic Layers limitation.

## 7. Save and hand off

1. Keep corrected Canva transactions in draft until the connector's required save approval is received.
2. One approval may cover all clearly identified variants.
3. Commit only variants that passed the real rendered QA.
4. Return:
   - one Canva edit link per ratio;
   - exact canvas dimensions;
   - a concise QA result;
   - editable text/image layer counts;
   - any parts that remain grouped.

## Selection guidance

Prefer this regeneration-first workflow for major orientation changes, illustration-heavy designs, or layouts that would need extensive manual recomposition. Prefer ordinary Canva resizing only when exact pixel-level continuity matters more than composition, or when the design is already a simple flat layout whose layers adapt cleanly.

--- references/qa.md ---

# Native-format and Magic Layers QA

Use this checklist on the actual rendered images, not only tool metadata.

## Generated reference image

- Canvas matches the requested ratio.
- Composition was rebuilt for the new orientation rather than padded, stretched, or cropped.
- The focal point and reading order are immediate.
- All required text is verbatim and readable at feed size.
- Lithuanian glyphs render correctly and consistently.
- People are intentionally framed and grounded; no accidental missing limbs or floating torsos.
- No text, clock, icon, card, calendar, or character collision exists.
- Empty space feels intentional and balanced.
- No invented copy, duplicate object, logo, or watermark appears.

## Magic Layers extraction

- Every intended text group is editable.
- OCR text matches the source character for character.
- Intentional line breaks and capitalization are preserved.
- Extracted image layers correspond to meaningful visual objects.
- Text boxes sit inside their intended containers.
- Icons do not cover text after extraction.
- Any grouped raster content is identified for the user.

## Post-edit Canva render

- Download and open the exact latest full-page preview.
- Confirm that the preview is current, not an earlier transaction image.
- Inspect headline hierarchy, margins, balance, contrast, and spacing.
- Check all corners and edges for clipping.
- Check actual glyph shapes, not only Unicode strings.
- Check short labels stay on one line.
- Check container-centred text is visually and geometrically centred.
- Check the subject crop looks intentional at the page edge.
- Check the result at thumbnail size for social-feed readability.

## Pass record

Record internally for each ratio:

```text
ratio and canvas:
source render inspected: yes/no
generated render inspected: yes/no
Magic Layers text audited: yes/no
latest Canva render downloaded and opened: yes/no
Lithuanian glyphs checked: yes/no
collisions and clipping checked: yes/no
orientation-specific composition checked: yes/no
editable text/image layers:
grouped raster elements:
result: pass/fail
```

Do not commit a `fail`.

Validate the Skill, install it, and confirm that it appears as `canva-new-generate`.

@canva-new-generate

Pagal šį originalą sukurk naujas redaguojamas „Canva“ versijas:
- 1:1;
- 16:9;
- 9:16.

Išlaikyk tą pačią pagrindinę mintį, tekstą, spalvas, personažą ir vizualinį stilių.
Kiekvieną formatą sugeneruok atskirai ir perkomponuok jo orientacijai – neapkirpk ir neištempk originalo.
Patikrink lietuviškas raides, teksto tilpimą, žmonių galūnes, elementų susidūrimus ir realų galutinį „Canva“ vaizdą.

@canva-new-generate

Iš šio vertikalaus 9:16 vizualo sukurk naują redaguojamą 16:9 „Canva“ versiją.
Tai turi būti nauja horizontali kompozicija, o ne apkirptas originalas.

@canva-new-generate

Naudok mano „Canva“ dizainą „[DIZAINO PAVADINIMAS]“ kaip vizualinį originalą.
Sukurk naujas 1:1 ir 16:9 versijas, kiekvieną sugeneruok atskirai ir paversk redaguojamu „Canva“ dizainu su „Magic Layers“.

Dar kartą atidaryk naujausią viso puslapio „Canva“ peržiūrą, ją realiai vizualiai įvertink ir išvardyk visus matomus trūkumus. Neišsaugok kaip baigto rezultato, kol jie neištaisyti.

Sukurta:
- 1:1 – 1080 × 1080 – QA praėjo – [Canva nuoroda]
- 16:9 – 1920 × 1080 – QA praėjo – [Canva nuoroda]
- 9:16 – 1080 × 1920 – QA praėjo – [Canva nuoroda]

Patikrinta: tekstas, lietuviškos raidės, elementų susidūrimai, kadravimas, viso puslapio „Canva“ peržiūra.