Photo Review Workflow

Photo Review Workflow

What This Skill Does

Use this skill for photo-shoot folders that contain original camera JPGs plus RAW/video sidecars. It preserves originals, organizes non-JPG files, evaluates only original JPG files with a fixed multimodal prompt, exports review tables, and optionally copies full-quality JPGs into rating folders for Finder browsing.

The bundled script is:

scripts/llm_photo_review.py

Default Workflow

1. Resolve the real photo directory. On macOS, Finder "图片" is usually ~/Pictures.
2. Create raw/ and move RAW extensions there: .RAF, .RAW, .CR2, .CR3, .NEF, .ARW, .DNG, .RW2, .ORF, .SRW.
3. Create 视频/ and move .MOV files there.
4. Score only original JPG files, normally DSCF*.JPG; ignore generated contact sheets, CSV/XLSX/PDF files, RAW, and videos.
5. Use one fixed prompt for every image. Do not tune the prompt per image or per scene.
6. Export:

• llm_aesthetic_review.csv
• llm_aesthetic_review.xlsx
• llm_aesthetic_review_top_200.pdf
• llm_aesthetic_review_summary.txt
• llm_aesthetic_review.status.json

1. Copy full-quality JPGs into rating folders only after scoring is complete.

Scoring Command

Use any Python 3 interpreter that has Pillow, ReportLab, and pandas/openpyxl-style dependencies installed:

PY=python3

For an OpenAI-compatible endpoint:

OPENAI_API_KEY="..." \
OPENAI_BASE_URL="https://your-openai-compatible-endpoint/v1" \
OPENAI_MODEL="gpt-5.4-mini" \
$PY scripts/llm_photo_review.py \
  --photos-dir /path/to/photos \
  --output-prefix llm_aesthetic_review \
  --max-side 384 \
  --jpeg-quality 70 \
  --workers 4 \
  --timeout 120 \
  --max-retries 3

Notes:

• Include /v1 in OPENAI_BASE_URL for OpenAI-compatible endpoints.
• If large image requests disconnect, reduce --max-side to 384 and --jpeg-quality to 70.
• The script supports resume by default: existing CSV rows are skipped.
• --no-resume forces a fresh run.
• --skip-errors skips files listed in the errors CSV.
• --pdf-only regenerates CSV/PDF artifacts from existing rows without API calls.

Fixed Prompt Intent

The script asks the model to act as a strict photography editor for street, travel documentary, and everyday images. Scores are aesthetic selection scores, not technical quality scores.

Core dimensions:

• subject and reason to look
• composition and visual form
• moment and story
• light and color
• editing/cropping potential
• technical floor, without over-rewarding sharp ordinary photos

Expected scoring is strict. A folder may have no A/B images; still use relative rank for selection.

Rating Folder Copy

After scoring, copy originals by file copy, not image re-encoding. Use a folder such as 按评分分类/ under the photo directory.

Recommended folder grouping:

按评分分类/
  C_可留/
  C_一般/
  D_一般/
  D_淘汰/
  E_淘汰/

Recommended filename:

999-score_scoreXXX_rankXXXX_GRADE_ORIGINAL.JPG

Example:

925_score074_rank0001_C_DSCF1568.JPG

999-score makes Finder name-sort place higher scores first while preserving the original filename for lookup.

Validation

Before final response, verify:

• source original JPG count equals scored row count
• no missing source files
• copied JPG count equals scored row count when rating folders are requested
• status JSON shows pending: 0
• unresolved errors are zero or clearly reported

Do not report API keys in final output or save them into skill files.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: KilimiaoSix
• Source: KilimiaoSix/agent-skills
• License: MIT

Install and usage instructions live in the source repository linked above.