skill存档

Teacher portrait assets

Configure the authoritative folder

Never infer the folder from a name such as 卡通. If project settings are empty, ask the user to identify the folder, then record it:

python scripts/workflow_guardrails.py teacher-folder-set `
  --settings <project-settings.json> --folder <user-specified-folder>

Only PNG files are selectable. Report matching JPG/ZIP files as unsupported rather than silently converting them.

Decide whether a portrait is needed

Inspect the chosen PSD first. For a generic PSD with a visible person slot, require a portrait from the configured authoritative folder. A contracted teacher-specific PSD variant is different: its exact teacher match satisfies the portrait requirement, and its built-in person remains unless the user explicitly requests a replacement. Never use a different teacher’s variant as the course poster. When the template contract records a user-approved base-derivation policy, a missing variant may be generated from the named base using this portrait workflow, but only the allowed portrait and teacher-name layers may change and the result must pass QA before registration.

For a generic PSD, record the requirement:

python scripts/workflow_guardrails.py teacher-asset-require `
  --manifest <manifest.json> --teacher "<teacher>" `
  --reason "poster has a person slot; authoritative portrait required" `
  --out <manifest.json>

Then plan candidates. An unneeded course is limited to a poster design with no visible person slot and records not_required_no_portrait_slot.

Name matching and multiple images

Match the teacher name without a trailing “老师”. Valid examples include 卡卡.png, 卡卡1.png, 卡卡_半身.png, and 卡卡形象二.png.

• zero matches: stop and request a PNG;
• one unique match: select it and show the filename;
• multiple matches: show every full filename and relative path, then ask the user to reply with the exact filename;
• duplicate identical filenames in subfolders: stop and ask the user to rename them uniquely because a filename alone cannot identify one.

Record an explicit selection:

python scripts/workflow_guardrails.py teacher-asset-select `
  --settings <project-settings.json> --manifest <manifest.json> `
  --teacher "<teacher>" --file-name "<exact filename.png>" `
  --out <manifest.json>

Transparency audit

Run alpha-audit, but treat its output as a screening signal only. A PNG may have an alpha channel while still containing a solid rectangular background. Inspect the image on checkerboard, white, and dark backgrounds.

The placed smart object or raster layer must contain only the person subject. Do not carry any original photo background, generated backdrop, floor, wall, frame, or rectangular matte into the final poster.

Background removal and confirmation

When the background is not transparent:

1. Use the image-editing workflow to isolate only the person.

2. Preserve identity, face, hair, clothing, pose, proportions, and canvas size. Do not beautify, redraw, add accessories, or alter the pose.

3. Save a temporary PNG outside the authoritative reference filename.

4. Show the cutout on checkerboard, white, and dark backgrounds. Inspect hair, fingers, clothing edges, holes, and residual color spill.

5. Ask the user for explicit confirmation.

6. After approval only, commit it:

   python scripts/workflow_guardrails.py cutout-commit `
     --settings <project-settings.json> `
     --original <reference.png> --processed <approved-cutout.png> `
     --user-approved

The helper preserves canvas size, requires real transparent pixels, creates a recoverable copy in _原图备份/, verifies hashes, and atomically replaces the original filename. Never delete the backup automatically.