NeoWeb3Nova/sticker-pack-maker-skill

Sticker Pack Maker

Create coherent sticker series rather than unrelated one-off illustrations. Preserve the reference character, make every scene semantically distinct, and deliver validated transparent PNG files plus an optional ZIP.

Output contract

• Produce one square PNG per scene.
• Preserve the character's face, hair, outfit, palette, outline weight, and rendering style.
• Keep requested speech-bubble text exact and legible.
• Use a thick white sticker border.
• Deliver RGBA images with fully transparent outer corners.
• Use stable numbered filenames such as 01_收到.png.
• Validate the count, dimensions, alpha channel, and corner transparency.

Workflow

1. Establish the pack brief

Derive sensible defaults instead of blocking when details are absent:

• Character: use the supplied reference image.
• Count: use the requested count; otherwise default to 20.
• Canvas: square, at least 1024×1024.
• Language: match the user's language.
• Style: match the reference image.
• Background: final transparent PNG.
• Delivery: individual PNG files and one ZIP.

Record any invariant shirt text, logo, accessories, or forbidden elements. Do not silently replace them.

2. Plan non-overlapping scenes

Create a numbered scene manifest before generation. Each scene needs:

• text: short bubble copy, usually 2–6 characters for Chinese.
• action: a visible pose or prop that communicates the meaning without text.
• emotion: a distinct facial expression.
• filename: stable numbered output name.

Read references/scene-planning.md when the user requests a themed pack or when more than 10 scenes are needed. Reuse a starter manifest from assets/scene-packs/ when it fits.

3. Generate chroma-key source images

Use the available image-generation tool with the character reference attached. Generate one image per call when exact text and consistency matter; small batches increase latency variance and make failures harder to recover.

Read references/prompt-patterns.md and build every prompt from the same character/style lock. Require:

• A perfectly flat, uniform #0000FF background.
• No blue elements in the character, props, border, or text.
• One character unless the scene explicitly requires a robot or secondary figure.
• Exact bubble text in quotation marks.
• No watermark, signature, unrelated text, duplicate limbs, or cropped sticker border.

Do not depend on native transparency from the generator. Chroma-key post-processing is deterministic and verifiable.

4. Convert to true transparency

Install the lightweight dependencies once:

python -m pip install -r skills/sticker-pack-maker/requirements.txt

Process, validate, and zip a source directory:

python skills/sticker-pack-maker/scripts/sticker_pipeline.py process \
  --input-dir ./generated-blue \
  --output-dir ./stickers-transparent \
  --zip ./stickers-transparent.zip \
  --expected-count 20 \
  --force

When generated filenames are random, provide a JSON manifest with source and filename fields:

python skills/sticker-pack-maker/scripts/sticker_pipeline.py process \
  --input-dir ./generated-blue \
  --output-dir ./stickers-transparent \
  --manifest ./manifest.json \
  --zip ./stickers.zip \
  --force

The script samples the border color, builds a soft alpha matte, removes chroma spill on fringe pixels, saves RGBA PNG files, validates them, and creates the archive.

5. Perform visual QA

Run deterministic validation:

python skills/sticker-pack-maker/scripts/sticker_pipeline.py validate \
  --input-dir ./stickers-transparent \
  --expected-count 20

Then visually inspect every image or at least a contact-sheet-sized view. Regenerate any image with:

• Incorrect or misspelled bubble text.
• Character drift or wrong shirt/logo text.
• Repeated scene composition.
• Extra fingers, limbs, people, or text.
• Blue subject elements damaged by chroma removal.
• Cropped white sticker borders.

Read references/quality-checklist.md for the full acceptance checklist.

6. Deliver and report

Return clickable paths to:

• The transparent PNG directory.
• The ZIP archive.
• One representative preview.

Report the count, dimensions, RGBA validation result, and any regenerated items. Never claim completion while image-generation calls are still queued.

Recovery rules

• If a generation call stalls, retain completed source images and retry only the missing scene.
• If exact text fails, simplify the composition and regenerate that image alone.
• If chroma removal damages blue subject details, regenerate with a non-blue palette or use a different key color and pass --key-color.
• If filenames no longer match scene order, build a manifest instead of relying on timestamps.
• Never delete original generated images during post-processing.

Resources

• scripts/sticker_pipeline.py: chroma removal, RGBA validation, renaming, and ZIP packaging.
• references/prompt-patterns.md: generation prompt locks and templates.
• references/scene-planning.md: scene diversity framework.
• references/quality-checklist.md: deterministic and visual QA.
• assets/scene-packs/: reusable AI, Web3, and workplace scene manifests.