geoff-ji/jiharn/document-processing

document-processing

Trigger

User sends a document (PDF / Word / Excel / PowerPoint / plain text) or asks to extract, summarise, or edit one. Also triggers when the user references a file path or URL ending in .pdf, .docx, .xlsx, .pptx, .txt, .md.

If the user mentions several files or links at once, treat this as a multi-document request and either process them in sequence or ask which one should go first.

Principle

Route each document kind to the right specialised tool. Never invent

content that is not in the document.

Decision: which backend?

1. Plain text / markdown (.txt, .md, .log):

→ use read_file (if local) or read_url (if remote).

1. Web-published HTML / article:

→ web-summary skill.

1. PDF:

→ if a PDF-aware MCP is installed (look for mcp__pdf__* or a plugin tool exposing pdf_read), use it. Otherwise propose installing an MCP via mcp-discovery (candidate examples: mcp-server-pdf, pdfplumber-mcp). Never fabricate content by guessing.

1. Word (`.docx`):

→ similar: mcp__docx__* or propose discovery.

1. Excel (`.xlsx`):

→ mcp__xlsx__* or code_execution with openpyxl for small local files, else propose MCP.

1. PowerPoint (`.pptx`):

→ mcp__pptx__* or propose MCP.

Inputs

• source: file path (inside workspace) or URL.
• format: explicit override when the extension lies (common for

URLs without suffix).

• operation: "extract" | "summarise" | "search" |

"convert".

• fields (optional): specific fields or pages to target.

Steps

1. Identify the document kind. If ambiguous, ask the user once.
2. Check the harness inventory for a matching MCP / plugin via

tool_search with the file type as query.

1. If no backend exists and the user wants to proceed:

• Route to mcp-discovery.
• Propose ONE candidate MCP with a one-line reason.
• Wait for confirmation before install.

1. If the user provided multiple documents, process them in the order given or ask which one to prioritise.
2. On success, extract / summarise / convert as requested.
3. Cite the page / sheet / slide numbers in the answer.

Verification

• Every fact in the reply can be pointed to a specific page / sheet /

cell.

• If the extraction returned partial results (rate limit, bad PDF,

password-protected), say so plainly — do not fill gaps with speculation.

• For conversions, the output file is actually created and its size

is non-zero.

Failure Signals

• The reply contains information that isn't in the source — stop,

re-read the actual document.

• "I cannot read PDFs" response with no mcp-discovery attempt — the

skill's job is to propose a path, not refuse.

• Installing multiple MCP servers for the same file kind — consolidate.

Output Format

文档：<source>  (<kind>, <size/pages>)
操作：<operation>
 
<内容摘要或提取，带页码 / 表格 / 列引用>
 
来源：p.X-Y / sheet:<name> / slide:<N>

If proposing an install:

这个文档需要 <kind> 支持，本地没有合适的工具。
建议装：<mcp name>  —  <一句话理由>
装吗？（yes/no）