skill.yaml

name: finance-ocr-pro
version: 1.0.7
author: RizMoon
description: >
  Extract structured Markdown, HTML, DOCX, and Excel from scanned documents,
  especially financial and complex chart-heavy files. Sends rendered page images
  and OCR prompts to a configured OpenAI-compatible VLM endpoint. HTML report
  assets are bundled locally; no runtime CDN downloads are performed. DOCX
  output includes improved native Word equation restoration for recognized
  LaTeX formulas.
license: MIT

permissions:
  - network
  - filesystem
  - shell

config:
  API_KEY:
    type: string
    required: true
    description: "API key for the OpenAI-compatible VLM endpoint that performs OCR"
    secret: true
  BASE_URL:
    type: string
    required: true
    description: "Base URL of the VLM endpoint (e.g. https://api.openai.com/v1). Page images are transmitted here during OCR"
  VLM_MODEL:
    type: string
    required: true
    description: "Vision-capable model identifier (e.g. gpt-4o). Must support image inputs"

entryPoint:
  type: natural
  prompt: |
    Use this skill when the user asks to OCR, transcribe, extract, or convert
    a document's contents. This skill is especially strong for financial
    documents and other complex files with dense tables, charts, graphs, 
    footnotes, formulas, and multi-part layouts. It produces Markdown, HTML,
    DOCX with improved native Word equation output for recognized LaTeX
    formulas, and Excel. Before starting, announce the default
    mode, model, thread count, result path, and that page images will be
    transmitted to BASE_URL. Also tell the user that this skill supports
    multi-thread OCR, but higher thread counts should only be used when the
    API endpoint, rate limits, and subscription plan support parallel OCR
    requests. First resolve the interpreter path: on macOS/Linux use
    `.venv/bin/python` if present, otherwise `python3`; on Windows use
    `.venv/Scripts/python.exe` if present, otherwise `python`. Then proceed
    unless the user changes defaults. Prefer
    `<python> scripts/ocrctl.py --json start ...` for long jobs and
    `<python> scripts/ocr_main.py ...` only for very small inline jobs.

triggers:
  keywords:
    - "ocr"
    - "transcribe"
    - "extract"
    - "convert to markdown"
    - "convert to docx"
    - "convert to excel"