Document Summarizer

Step 2: Check Dependencies

python3 "$SKILL_DIR/scripts/check_dependencies.py"

• Exit 0: all good. Exit 1: packages were installed (proceed). Exit 2: failed (report to user).

Step 3: Chunk the Document(s)

Determine the work directory based on input type:

• Single file: WORK_DIR="{parent_dir}/{filename_without_ext}_summary_work"
• Directory: WORK_DIR="{directory_path}/_summary_work"

This places chunks alongside the source so users can review them.

mkdir -p "$WORK_DIR"

python3 "$SKILL_DIR/scripts/chunk_document.py" \
  "<file_or_directory_path>" \
  "$WORK_DIR" \
  --max-tokens 4000 \
  --overlap 200

The script accepts either a single file or a directory. Read $WORK_DIR/metadata.json to determine the mode.

metadata.json mode field:

• "single_file": one document was processed. Chunks are in chunks array.
• "multi_file": a directory was processed. Each file is in the files array, each with its own chunks sub-array.

Step 4: Determine Strategy

Read metadata.json. Count total chunks:

• Single file: num_chunks field
• Multi file: total_chunks field

Small job (1-3 total chunks): Summarize directly — no team needed.

• Read each chunk file sequentially
• Skip to Step 6 output format

Medium/large job (4+ total chunks): Create an agent team.

• Calculate agent count: min(8, max(2, total_chunks // 2))
• Proceed to Step 5

Step 5: Agent Team Coordination

5a. Create the Team

TeamCreate: team_name="doc-summary", description="Summarizing <name>"

5b. Spawn Summarizer Agents

Divide chunks evenly across agents. Keep contiguous chunks together. For multi-file mode, keep chunks from the same file together when possible.

For each agent, spawn via Task tool with subagent_type: "general-purpose" and this prompt:

You are summarizing sections of a large document.

Read these chunk files and write a summary for each:
{list of absolute chunk file paths, e.g. $WORK_DIR/chunks/chunk_001.txt}

For context, here is the chunk metadata:
{chunk entries from metadata.json for assigned chunks}

Write your output to: {WORK_DIR}/summaries/section_{agent_number}.md

Use this format for your output file:

## {heading from chunk metadata}
**Source file**: {filename, if multi-file mode}
**Pages**: {start_page}-{end_page} (omit if pages are 0)

### Summary
[2-4 paragraphs summarizing the content]

### Key Points
- [Important point 1]
- [Important point 2]

### Notable Details
- [Specific data, statistics, quotes, or references worth preserving]

---

Repeat the above for each chunk you are assigned.
After writing the file, confirm completion.

Launch all agents in parallel (multiple Task tool calls in one message).

5c. Collect Results

After all agents complete:

1. Read all summary files: {WORK_DIR}/summaries/section_*.md
2. Read metadata.json for structure
3. Proceed to Step 6

5d. Clean Up

After producing the final output:

• Send shutdown_request to all agents
• TeamDelete to clean up

Step 6: Produce Final Output

The final deliverable is a .docx file placed in the same folder as the original document(s).

Output file naming:

• Single file: {original_filename_without_ext}_summary.docx
• Directory: Summary_{dirname}.docx

How to generate the file: Use the npm docx package to generate the .docx file from a Node.js script.

Also write a plain-text copy to {WORK_DIR}/final_summary.md for reference.

Document structure requirements (for the .docx):

• Title page with document name, date, page/token counts
• Table of Contents using HeadingLevel styles
• Header with document title, footer with page numbers
• Professional styling: Arial font, proper heading hierarchy, consistent spacing
• Tables for structured data (file listings, effective dates, etc.)
• Proper bullet lists (using numbering config, not unicode)
• Page breaks between major sections

Content template for SINGLE FILE mode:

The .docx should contain:

1. Title: "Document Summary: {filename}"
2. Metadata block: Source path, pages, token count, sections processed
3. Executive Summary (Heading 1): 2-3 paragraphs covering what the document is about, main conclusions, intended audience, and key takeaways
4. Document Structure (Heading 1): Numbered outline with section headings and page ranges
5. Section Summaries (Heading 1): For each section:
   • Section heading (Heading 2) with page range
   • Summary paragraphs
   • Key Points as bullet list
   • Notable Details as bullet list
6. Key Findings and Takeaways (Heading 1): Numbered list of the most important findings
7. Notable Data and References (Heading 1): Key statistics, dates, figures, citations, named entities

Content template for MULTI-FILE mode:

The .docx should contain:

1. Title: "Document Collection Summary"
2. Metadata block: Source directory, file count, total tokens, total chunks
3. Files Processed table: columns for #, Filename, Type, Pages, Tokens, Chunks
4. Executive Summary (Heading 1): 2-3 paragraphs synthesizing themes across ALL documents
5. Per-Document Summaries (Heading 1): For each document:
   • Document name (Heading 2) with type, pages, tokens
   • Structure outline
   • Summary paragraphs
   • Key Points as bullet list
6. Cross-Document Findings (Heading 1): Themes and patterns that span multiple documents
7. Notable Data and References (Heading 1): Key statistics, dates, figures across all documents

Accuracy and QA (Required)

Anti-hallucination rules (include in ALL subagent prompts):

• Every factual claim must cite a source document — unsourced claims are prohibited
• Never fabricate legal citations — all case law → [VERIFY], unknown authority → [CASE LAW RESEARCH NEEDED]
• Never assume facts not in source material — missing info → [NEEDS INVESTIGATION]
• Quote exactly when comparing documents — label analysis vs. facts distinctly

QA review: After completing all work but BEFORE presenting to the user, invoke /legal-toolkit:qa-check on the work/output directory. Do not skip this step.

Error Handling

• Path not found: Ask user to verify the path
• Unsupported format: Supported types are .pdf, .docx, .txt, .md
• Empty directory: No supported files found — ask user to check the folder
• Empty extraction: File may be scanned/image-only. Delegate OCR to a subagent: launch an Agent (subagent_type: "general-purpose") with prompt: "Run /legal-toolkit:extract-text on {file_path} and write the extracted text to $WORK_DIR/{filename}_ocr.txt." Re-run chunking on the OCR output.
• Agent failure: Read the unprocessed chunks directly and summarize yourself
• Script not found: Verify the skill is installed (ls $SKILL_DIR/scripts/)