Sn Da Non Spreadsheet AnalysisSAFE

Modular SenseNova skills for building AI-powered office assistants and productivity workflows

The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.

---
name: sn-da-non-spreadsheet-analysis
description: "Word / PDF / PPT 文档解析与数据分析引擎。覆盖三类文件格式的全量提取、表格数值化、图表理解与跨文档汇总分析。**遇到以下任一情况就主动使用本 skill**:1用户上传或指定了 .docx / .doc / .pdf / .pptx / .ppt 文件并要求分析、提取或统计其中内容;2用户出现触发词:Word分析 / PDF解析 / PPT提取 / 文档分析 / 报告解析 / 幻灯片分析 / 发票提取 / 合同分析 / 文档统计 / 错别字 / 语病 / 字号检查 / 简历分析 / 多文档对比;3任务涉及从文档中提取表格、数值、图表、格式(颜色/高亮/字号)、组织架构、时间线等结构化信息。仅不用于:Excel/CSV 数据分析(使用 sn-da-excel-workflow)、纯图片分析(使用 sn-da-image-caption)。"
---

# Document Analysis Skill — Word / PDF / PPT

End-to-end workflow for Word, PDF, and PPT document parsing. Each format has
specific parsing pitfalls — follow the format-specific sub-skill exactly.

---

## Workflow

### Step 0 — Identify file type and input scope

```python
import os

input_path = "/mnt/data/..."  # from user

# Detect single file vs directory (multi-file scenario)
if os.path.isdir(input_path):
    all_files = [
        os.path.join(input_path, f)
        for f in os.listdir(input_path)
        if f.lower().endswith(('.docx', '.doc', '.pdf', '.pptx', '.ppt'))
    ]
    print(f"Found {len(all_files)} documents: {all_files}")
else:
    all_files = [input_path]

# Route by extension
ext = os.path.splitext(all_files[0])[-1].lower()
print(f"File type: {ext}")
```

> **Critical rule**: When `input_path` is a directory OR the user says "这些文件" / "所有文档",
> process **every file** and aggregate. Never stop at the first file.

---

### Step 1 — Load sub-skill by format

| Extension | Sub-skill to load |
|-----------|------------------|
| `.docx` / `.doc` | `capability/word-analysis/SKILL.md` |
| `.pdf` | `capability/pdf-analysis/SKILL.md` |
| `.pptx` / `.ppt` | `capability/ppt-analysis/SKILL.md` |

```
read_file(path="<skills_root>/sn-da-non-spreadsheet-analysis/capability/<format>-analysis/SKILL.md")
```

Load **only the sub-skill you need** — do not load all three at once.

---

### Step 2 — Parse and extract

Follow the sub-skill's extraction pattern. For all formats:

- **Full scan**: iterate all pages/slides/paragraphs — never stop early
- **Table extraction**: get every table, not just the first one
- **Image/chart detection**: if a page/slide yields no text, treat it as image-based and call `caption.py`

---

### Step 3 — Answer with verification

After extracting data, verify before answering:

```python
# For count/statistics questions: spot-check 3-5 items
sample = result_list[:3]
print(f"Sample check: {sample}")
print(f"Total count: {len(result_list)}")

# For numeric calculations: print intermediate values
print(f"Max={max_val}, Min={min_val}, Range={max_val - min_val}")

# For unit-sensitive answers: always include the unit
print(f"Answer: {value} {unit}")  # e.g., "475 千港元" not just "475"
```

---

## Universal Rules

### MUST DO
- **Always iterate all pages/slides/paragraphs** — `for page in doc`, `for slide in prs.slides`, `for para in doc.paragraphs`
- **When input is a directory**: collect and process all matching files, then aggregate results
- **For scanned PDFs**: detect empty text → call `caption.py` for OCR
- **For image-only slides**: text extraction returns empty → render slide as PNG → call `caption.py`
- **For calculations**: show intermediate values; confirm unit matches the question

### NEVER DO
- Do NOT use `pytesseract` or `easyocr` as primary OCR — they are not installed; use `caption.py`
- Do NOT use PIL pixel analysis to infer chart values — use vision model caption instead
- Do NOT stop at the first file, first page, or first table
- Do NOT guess content from filenames — always parse the actual file
- Do NOT output percentage when the question asks for absolute value (and vice versa)

---

## Caption Script (for image/chart content in any document)

When a page, slide, or embedded image needs vision understanding, load the
`sn-da-image-caption` skill first, then use its `scripts/caption.py`:

```
read_file(path="<skills_root>/sn-da-image-caption/SKILL.md")
```

```python
import subprocess, json

CAPTION = "/path/to/skills/sn-da-image-caption/scripts/caption.py"

def caption_image(image_path, prompt=None):
    cmd = ["python3", CAPTION, image_path, "--json"]
    if prompt:
        cmd += ["--prompt", prompt]
    result = subprocess.run(cmd, capture_output=True, text=True, timeout=60)
    if result.returncode != 0:
        raise RuntimeError(f"caption failed: {result.stderr[:200]}")
    return json.loads(result.stdout)["description"]

# Example prompts by content type:
# Table:  "提取表格所有内容,Markdown 表格格式,保持行列结构,数值不四舍五入。"
# Chart:  "提取图表标题、坐标轴标签、每个数据点的数值。Markdown 表格输出。"
# Diagram: "描述所有节点和连接关系。"
```

---

## Available sub-skills

```
sn-da-non-spreadsheet-analysis/capability/word-analysis/SKILL.md   — .docx/.doc
sn-da-non-spreadsheet-analysis/capability/pdf-analysis/SKILL.md    — .pdf
sn-da-non-spreadsheet-analysis/capability/ppt-analysis/SKILL.md    — .pptx/.ppt
```

SAFEgrade B · trust 89/100 Nothing in the source contradicts what it says it does. Grade A is reserved for packages that have also passed the behavioural sandbox.

What the source does

Filesystem

none-observed

Network

none-observed

Shell

none-observed

Dependencies

pinned

Secrets in source

none-found

Findings (0)

No findings outside the package's declared scope.

Gates applied: no_behavioural_pass.