做什么
> 翻译:
技能库 智客分类:Agent 工作流 smart-ocr
> 翻译:
官方网址:skills.sh
先看中文介绍;官方 description 原文单独保留,不改写 SKILL.md。
> 翻译:
官方 description 未单独写出 Use when。按规范,代理会在用户任务与这段 description 的关键词匹配时激活本技能。
按 Agent Skills 渐进披露:启动时只加载 name 与 description(约 100 token);任务匹配后才读入整份 SKILL.md 正文;scripts/、references/、assets/ 仅在需要时再读。 本文件正文结构:Smart OCR Skill、Overview、How to Use、Domain Knowledge、PaddleOCR Fundamentals、Initialize OCR engine。 其中含规范建议的小节:边界情况。
文件分析:这是一份仅含 SKILL.md 的指令型技能,代理激活后整份正文进入上下文。
>
Smart OCR SkillOverviewHow to UseDomain KnowledgePaddleOCR FundamentalsInitialize OCR engineRun OCR on imageResult structure: [[box, (text, confidence)], ...]Supported LanguagesCommon language codesUse specific languageConfiguration Options
· 许可:MIT
来源分类:skills.sh agent-skill
namesmart-ocrdescription具体调用语法与可用工具以目标 Agent 客户端为准。 查看调用机制说明 ↗
先选择目标 Agent 和安装范围,保留技能包的附属文件,安装后检查客户端能否发现该技能。
复制安装指令给支持 Agent Skills 的代理,确认其中的目标目录与客户端匹配。
把 Agent Skill「smart-ocr」安装到我的项目:SKILL.md 原文与官方 description 见 https://zicq.com/zh/skills/skl-edb05103082ecd58-%E8%81%AA%E6%98%8E%E7%9A%84-Ocr.html 请存为 .cursor/skills/smart-ocr/SKILL.md 或 .claude/skills/smart-ocr/SKILL.md,frontmatter 的 name 与 description 保持原样,不要改写。
需要 Node.js 与 npx。先查看仓库技能列表,确认实际名称。
npx skills add 'https://github.com/claude-office-skills/skills' --list
npx skills add 'https://github.com/claude-office-skills/skills' --skill 'smart-ocr'
CLI 会交互选择目标 Agent,默认安装到项目;用户级安装使用 -g。先通过查看命令核对仓库内容,再用 npx skills list 检查已安装技能。
name: smart-ocr description: ">" version: "1.0" author: claude-office-skills license: MIT
category: parsing tags:
models: recommended: - claude-sonnet-4 - claude-opus-4 compatible: - claude-3-5-sonnet - gpt-4 - gpt-4o
mcp: server: office-mcp tools: - extract_text_from_pdf
capabilities:
languages:
This skill enables intelligent text extraction from images and scanned documents using PaddleOCR - a leading OCR engine supporting 100+ languages. Extract text from photos, screenshots, scanned PDFs, and handwritten documents with high accuracy.
Example prompts:
from paddleocr import PaddleOCR
# Initialize OCR engine
ocr = PaddleOCR(use_angle_cls=True, lang='en')
# Run OCR on image
result = ocr.ocr('image.png', cls=True)
# Result structure: [[box, (text, confidence)], ...]
for line in result[0]:
box = line[0] # [[x1,y1], [x2,y2], [x3,y3], [x4,y4]]
text = line[1][0] # Extracted text
conf = line[1][1] # Confidence score
print(f"{text} ({conf:.2f})")
# Common language codes
languages = {
'en': 'English',
'ch': 'Chinese (Simplified)',
'cht': 'Chinese (Traditional)',
'japan': 'Japanese',
'korean': 'Korean',
'french': 'French',
'german': 'German',
'spanish': 'Spanish',
'russian': 'Russian',
'arabic': 'Arabic',
'hindi': 'Hindi',
'vi': 'Vietnamese',
'th': 'Thai',
# ... 100+ languages supported
}
# Use specific language
ocr = PaddleOCR(lang='ch') # Chinese
ocr = PaddleOCR(lang='japan') # Japanese
ocr = PaddleOCR(lang='multilingual') # Auto-detect
from paddleocr import PaddleOCR
ocr = PaddleOCR(
# Detection settings
det_model_dir=None, # Custom detection model
det_limit_side_len=960, # Max side length for detection
det_db_thresh=0.3, # Binarization threshold
det_db_box_thresh=0.5, # Box score threshold
# Recognition settings
rec_model_dir=None, # Custom recognition model
rec_char_dict_path=None, # Custom character dictionary
# Angle classification
use_angle_cls=True, # Enable angle classification
cls_model_dir=None, # Custom classification model
# Language
lang='en', # Language code
# Performance
use_gpu=True, # Use GPU if available
gpu_mem=500, # GPU memory limit (MB)
enable_mkldnn=True, # CPU optimization
# Output
show_log=False, # Suppress logs
)
# Single image
result = ocr.ocr('image.png')
# Multiple images
images = ['img1.png', 'img2.png', 'img3.png']
for img in images:
result = ocr.ocr(img)
process_result(result)
from pdf2image import convert_from_path
def ocr_pdf(pdf_path):
"""OCR a scanned PDF."""
# Convert PDF pages to images
images = convert_from_path(pdf_path)
all_text = []
for i, img in enumerate(images):
# Save temp image
temp_path = f'temp_page_{i}.png'
img.save(temp_path)
# OCR the image
result = ocr.ocr(temp_path)
# Extract text
page_text = '\n'.join([line[1][0] for line in result[0]])
all_text.append(f"--- Page {i+1} ---\n{page_text}")
os.remove(temp_path)
return '\n\n'.join(all_text)
import requests
from io import BytesIO
# From URL
response = requests.get('https://example.com/image.png')
result = ocr.ocr(BytesIO(response.content))
# From bytes
with open('image.png', 'rb') as f:
img_bytes = f.read()
result = ocr.ocr(BytesIO(img_bytes))
def process_ocr_result(result):
"""Process OCR result into structured data."""
lines = []
for line in result[0]:
box = line[0]
text = line[1][0]
confidence = line[1][1]
# Calculate bounding box
x_coords = [p[0] for p in box]
y_coords = [p[1] for p in box]
lines.append({
'text': text,
'confidence': confidence,
'bbox': {
'left': min(x_coords),
'top': min(y_coords),
'right': max(x_coords),
'bottom': max(y_coords),
},
'raw_box': box
})
return lines
# Sort by position (top to bottom, left to right)
def sort_by_position(lines):
return sorted(lines, key=lambda x: (x['bbox']['top'], x['bbox']['left']))
def reconstruct_layout(result, line_threshold=10):
"""Reconstruct text layout from OCR results."""
lines = process_ocr_result(result)
lines = sort_by_position(lines)
# Group into logical lines
text_lines = []
current_line = []
current_y = None
for line in lines:
y = line['bbox']['top']
if current_y is None or abs(y - current_y) < line_threshold:
current_line.append(line)
current_y = y
else:
# New line
text_lines.append(' '.join([l['text'] for l in current_line]))
current_line = [line]
current_y = y
# Add last line
if current_line:
text_lines.append(' '.join([l['text'] for l in current_line]))
return '\n'.join(text_lines)
from PIL import Image, ImageEnhance, ImageFilter
def preprocess_image(image_path):
"""Preprocess image for better OCR."""
img = Image.open(image_path)
# Convert to grayscale
img = img.convert('L')
# Enhance contrast
enhancer = ImageEnhance.Contrast(img)
img = enhancer.enhance(2.0)
# Sharpen
img = img.filter(ImageFilter.SHARPEN)
# Save preprocessed
preprocessed_path = 'preprocessed.png'
img.save(preprocessed_path)
return preprocessed_path
from tqdm import tqdm
from concurrent.futures import ThreadPoolExecutor
def batch_ocr(image_paths, max_workers=4):
"""OCR multiple images in parallel."""
results = {}
def process_single(img_path):
result = ocr.ocr(img_path)
return img_path, result
with ThreadPoolExecutor(max_workers=max_workers) as executor:
futures = [executor.submit(process_single, p) for p in image_paths]
for future in tqdm(futures, desc="Processing OCR"):
path, result = future.result()
results[path] = result
return results
from paddleocr import PaddleOCR
import re
def read_business_card(image_path):
"""Extract contact info from business card."""
ocr = PaddleOCR(use_angle_cls=True, lang='en')
result = ocr.ocr(image_path)
# Extract all text
all_text = []
for line in result[0]:
all_text.append(line[1][0])
full_text = '\n'.join(all_text)
# Parse contact info
contact = {
'name': None,
'email': None,
'phone': None,
'company': None,
'title': None,
'raw_text': full_text
}
# Email pattern
email_match = re.search(r'[\w\.-]+@[\w\.-]+\.\w+', full_text)
if email_match:
contact['email'] = email_match.group()
# Phone pattern
phone_match = re.search(r'[\+\d][\d\s\-\(\)]{8,}', full_text)
if phone_match:
contact['phone'] = phone_match.group().strip()
# Name is usually the largest/first text
if all_text:
contact['name'] = all_text[0]
return contact
card_info = read_business_card('business_card.jpg')
print(f"Name: {card_info['name']}")
print(f"Email: {card_info['email']}")
print(f"Phone: {card_info['phone']}")
from paddleocr import PaddleOCR
import re
def scan_receipt(image_path):
"""Extract items and total from receipt."""
ocr = PaddleOCR(use_angle_cls=True, lang='en')
result = ocr.ocr(image_path)
lines = []
for line in result[0]:
text = line[1][0]
y_pos = line[0][0][1]
lines.append({'text': text, 'y': y_pos})
# Sort by vertical position
lines.sort(key=lambda x: x['y'])
receipt = {
'items': [],
'subtotal': None,
'tax': None,
'total': None
}
for line in lines:
text = line['text']
# Look for total
if 'total' in text.lower():
amount = re.search(r'\$?([\d,]+\.?\d*)', text)
if amount:
if 'sub' in text.lower():
receipt['subtotal'] = float(amount.group(1).replace(',', ''))
else:
receipt['total'] = float(amount.group(1).replace(',', ''))
# Look for tax
elif 'tax' in text.lower():
amount = re.search(r'\$?([\d,]+\.?\d*)', text)
if amount:
receipt['tax'] = float(amount.group(1).replace(',', ''))
# Look for items (line with price)
else:
item_match = re.search(r'(.+?)\s+\$?([\d,]+\.?\d+)$', text)
if item_match:
receipt['items'].append({
'name': item_match.group(1).strip(),
'price': float(item_match.group(2).replace(',', ''))
})
return receipt
receipt_data = scan_receipt('receipt.jpg')
print(f"Items: {len(receipt_data['items'])}")
print(f"Total: ${receipt_data['total']}")
from paddleocr import PaddleOCR
def ocr_multilingual(image_path, languages=['en', 'ch']):
"""OCR document with multiple languages."""
all_results = {}
for lang in languages:
ocr = PaddleOCR(use_angle_cls=True, lang=lang)
result = ocr.ocr(image_path)
texts = []
for line in result[0]:
texts.append({
'text': line[1][0],
'confidence': line[1][1]
})
all_results[lang] = texts
# Merge results, keeping highest confidence
merged = {}
for lang, texts in all_results.items():
for item in texts:
text = item['text']
conf = item['confidence']
if text not in merged or merged[text]['confidence'] < conf:
merged[text] = {'confidence': conf, 'language': lang}
return merged
result = ocr_multilingual('bilingual_document.png')
for text, info in result.items():
print(f"[{info['language']}] {text} ({info['confidence']:.2f})")
# CPU version
pip install paddlepaddle paddleocr
# GPU version (CUDA 11.x)
pip install paddlepaddle-gpu paddleocr
# Additional dependencies
pip install pdf2image Pillow
Agent 工作流
创造有效技能指南。 当用户想创造出新的技能(或更新现有的技能),以专业知识,工作流程,或工具集成来扩展克洛德的能力时,应该使用这种技能.
Agent 工作流
使用ClawdHub CLI搜索,安装,更新并发布从taladhub.com的代理技能. 需要获取苍蝇上的新技能时使用,将安装的技能同步到最新版本或特定版本,或者发布 npm-instainddhub CLI 的新/更新的技能文件夹.
Agent 工作流
管弦乐团多代理团队,任务设定周期,交接协议,审查工作流程. 使用时间: (1)建立2+特派员队伍,具有不同专业,(2)确定任务路线和生命周期(收录框_ spec_建设_审查_完成),(3)在特派员之间制定交接协议,(4)建立审查和质量关口,(5)管理特派员之间的交流和文物共享.
Agent 工作流
Spec-first,TDD,子代理驱动的软件开发工作流程. 当:(1)构建任何新功能或应用——触发脑暴_计划_子代理执行回路,(2)调试出一个bug或测试失败——触发系统性的根起过程,(3)用户说"让我们构建","帮助我计划","我想添加X",或"这个被打破",(4)完成一个功…