resume

name	resume
description	Synthesize long documents (PDF, Markdown, URL) into concise summaries in English. Uses hierarchical table-of-contents chunking with parallel sub-agents to process hundreds of pages without context loss. Use when user wants to summarize a long document, says "summarize this report", "give me a synthesis of this document", "what does this PDF say", or mentions /resume. Triggers on any request to summarize, synthesize, or extract key points from reports, white papers, or long-form content.
argument_hint	<file|url> [--words N] [--detailed|--concise] [--lang fr]

resume — Long Document Synthesis

Hierarchical synthesis with parallel sub-agents. Strategy B (table of contents) with fallback A (fixed-size chunks).

Parameters

Parameter	Effect
<source>	PDF/Markdown path or URL
--words N	Target length in words (default: ~500)
--detailed	Long summary (~1000 words)
--concise	Short summary (~250 words)
--lang fr	Output language (default: en)

Workflow

1. Get the text

• HTML URL → webfetch with extract_main: true
• PDF URL → curl -L <url> -o /tmp/resume-input.pdf
• Local PDF → python3 .claude/skills/resume/scripts/extract_pdf.py <pdf> > /tmp/resume-extracted.json
• Markdown → direct read with Read tool

The script outputs a JSON: {text, toc: [{title, page, level}], pages}.

2. Detect structure and chunk

If table of contents available (Strategy B) :

• Group ToC entries into chunks of ~12K tokens max
• Each chunk = one or more contiguous sections
• Keep titles as chunk headers

If no ToC (Strategy A, automatic fallback) :

• Split text into blocks of ~12K tokens
• 200 token overlap between blocks

If < 2 pages → direct summary, no chunking.

3. Summarize chunks in parallel

Launch one sub-agent per chunk with this prompt:

Summarize the following text in {language}. Capture:
- Main facts and findings
- Key numbers and data points
- Conclusions and recommendations
- Essential context

Length: {1-2% of chunk size, max 800 words}.

Collect all summaries. Use the cheapest model capable of proper summarization for chunks.

4. Final synthesis

With all chunk summaries assembled, produce the final synthesis:

• Language: English by default, or as requested via --lang
• Length: per --words, --detailed, or --concise
• Format: freeform, structured naturally based on content
• Prioritize: conclusions, implications, controversies if present

Use the primary model for this step.

5. Write output

Create the file in Resources/ with this frontmatter:

type: summary
source: "<url or path>"
source_type: <pdf|markdown|url>
lang: <en|fr>
tags: [summary]
created: <YYYY-MM-DD>

Display the summary in chat, then confirm the path: Summary saved: Resources/<slug>.md.

Error handling

Case	Action
PDF with no text (scan)	Message: "This PDF is a scan without OCR. Use an OCR tool first."
URL unreachable	Display HTTP code, stop
Document without ToC	Silent fallback to Strategy A
Document < 2 pages	Direct summary, no chunking
pymupdf missing	Message: pip install pymupdf
Chunk > 12K tokens	Recursive re-chunking of that chunk

Notes

• Extracted text → /tmp/, cleaned up after use
• Always include the exact source in the frontmatter
• File slug = document title in kebab-case
• For URLs, webfetch with prefer_llms_txt: auto by default
• Requires pymupdf for PDF extraction and summarize CLI for content extraction