AgentSkillsCN

image-transcription

将螺旋笔记本的照片与手写笔记中的羊群记录进行转录。从笔记本页面的图像中提取治疗数据、测量结果、圈舍分配,以及羊只的身份信息。

SKILL.md
--- frontmatter
name: image-transcription
description: Transcribe sheep records from spiral notebook photos and handwritten notes. Handle extraction of treatment data, measurements, pen assignments, and sheep identification from notebook page images.

Image Transcription for Sheep Records

Core Principles

  1. Accuracy Over Speed — One correct record is worth more than ten guessed ones
  2. Source Fidelity — Record exactly what the image shows, then normalize
  3. Honesty About Uncertainty — Use [UNCLEAR] for anything you can't confidently read

Pre-Transcription Checklist

  • Image is from data/processed/ directory (≤1800px)
  • Identify the type: spiral notebook note, handwritten card, sheep photo, ear tag close-up
  • Note the date if visible (phone notes show timestamps)

Image Types

Phone App Notes (PNG files, IMG_8616–8643)

  • Treatment records with FAMACHA scores
  • Measurement records (height, girth)
  • Pen assignment lists
  • Sick sheep records with dates
  • Health notes and vaccination records

Handwritten Notebook Pages (JPG files, IMG_8560–8615)

  • Detailed sheep check records
  • Close-up photos of individual sheep with ear tags
  • Breed identification notes
  • Lambing records

Transcription Format

For each sheep data point found, record:

  • Source image: IMG_####
  • Data type: treatment | measurement | pen_assignment | health_note | lambing | identification
  • Date: if visible in the image
  • Content: exact transcription of the text/data
  • Normalized: cleaned version for the database

Struck-Through Text

Text with strikethrough in the notebook generally means:

  • The animal is deceased or sold
  • The treatment was not given (e.g., "no treat" struck through)
  • The entry was corrected elsewhere

Always note strikethrough and interpret carefully.

MUST DO

  • Use data/processed/*.jpeg versions only
  • Note which image each data point comes from
  • Preserve original spellings in notes field
  • Mark uncertain readings as [UNCLEAR]

MUST NOT

  • Read original images (>2000px)
  • Guess unclear handwriting
  • Invent data not visible in the image
  • Discard information because it seems contradictory (record it and flag it)