multimodal-evidence

Agent Skills › odysseus-dev/odysseus › multimodal-evidence

multimodal-evidence

GitHub

从图像、文档和视频中提取并验证证据，避免冗余检查。通过结构化定位、原生解析及针对性复核，确保事实输出可追溯且覆盖所有请求项。

resources/skills/media/multimodal-evidence/SKILL.md odysseus-dev/odysseus

Trigger Scenarios

需要从图片、文档或视频中提取具体事实数据需要验证多模态内容中的文本、数值或视觉关系

Install

npx skills add odysseus-dev/odysseus --skill multimodal-evidence -g -y

More Options

Non-standard path

npx skills add https://github.com/odysseus-dev/odysseus/tree/dev/resources/skills/media/multimodal-evidence -g -y

Use without installing

npx skills use odysseus-dev/odysseus@multimodal-evidence

指定 Agent (Claude Code)

npx skills add odysseus-dev/odysseus --skill multimodal-evidence -a claude-code -g -y

安装 repo 全部 skill

npx skills add odysseus-dev/odysseus --all -g -y

预览 repo 内 skill

npx skills add odysseus-dev/odysseus --list

SKILL.md

Frontmatter

{
    "name": "multimodal-evidence",
    "tags": [
        "image",
        "video",
        "document",
        "evidence",
        "ocr"
    ],
    "owner": "",
    "source": "builtin",
    "status": "published",
    "created": "2026-08-30T00:00:00Z",
    "version": "1.0.1",
    "category": "media",
    "confidence": 1,
    "description": "Extract and verify evidence from images, documents, and video without redundant inspection"
}

When to Use

Use when the answer or requested artifact depends on visual, temporal, tabular, or textual evidence contained in images, documents, or video.

Procedure

Identify the evidence required: objects, text, values, ordering, timestamps, labels, or visual relationships.
Inspect the whole input or a broad representative sample first to establish structure and likely evidence locations.
Narrow to relevant pages, frames, regions, or time intervals and record observations with their locations.
Use the format's native parser for exact text and numbers: for example python-docx or ZIP/XML inspection for DOCX, pdftotext or a PDF library for PDF, spreadsheet readers for XLSX, and OCR only when the source is image-based. Do not search binary office files with plain grep or cat.
Resolve conflicts with one targeted reinspection at better scale or a nearby frame rather than repeating the same crop.
Build the answer or artifact from the evidence ledger and perform a final coverage check against every requested item.

Pitfalls

Do not infer unseen content from filenames, surrounding text, or a single thumbnail.
Do not repeatedly inspect nearly identical regions without a new hypothesis.
Do not trust OCR blindly for small labels, punctuation, or numeric values.
Do not finalize before checking that every requested item has supporting evidence.