research-docs（研究文档）

Research Docs — Document Q&A with Visual Citations

Parse documents with LiteParse, answer a question using the parsed text, and generate an HTML report with source citations highlighted on page images.

Arguments

$ARGUMENTS should contain:

• First argument ($0): Path to the data directory containing documents
• Remaining arguments: The question to answer

If either is missing, ask the user to provide them.

Step 1 — Parse Documents

IMPORTANT: Always use the bundled Python script below for parsing. Do NOT call lit or liteparse CLI commands directly — use only generate_report.py.

Run the bundled parse script to extract text and bounding boxes from all supported files:

python "${CLAUDE_SKILL_DIR}/scripts/generate_report.py" \
    --skill-dir "${CLAUDE_SKILL_DIR}" \
    --dir "$0" \
    --parse-only \
    --output /tmp/research_docs_parsed.json

This discovers and parses all supported files in the directory:

• LiteParse formats: PDF, DOCX, PPTX, XLSX, images (up to 50 files)
• Plaintext:.txt,.md,.rst (read directly)

The output is a JSON file with parsed text and bounding box coordinates for each page.

If the directory has more than 50 files and the user's question targets a specific document not in the first 50, re-run with a narrower --dir pointing to a subdirectory, or ask the user which files to focus on.

Step 2 — Read Parsed Content

Read /tmp/research_docs_parsed.json using the Read tool. Focus on:

• Each file's name and type
• For LiteParse files: each page's text field (skip raw textItems — those are for bounding box rendering)
• For plaintext files: the text field
• The summary object for total counts

Build a mental model of all document content before answering.

Step 3 — Answer with Citations

Using the parsed text as context, answer the user's question. Write your response as a JSON file:

cat > /tmp/research_docs_answer.json << 'ANSWER_EOF'
{
  "question": "<the user's question>",
  "answer": "<your answer in markdown with [N] citation markers>",
  "citations": [
    {
      "file": "<filename e.g. report.pdf>",
      "page": <1-indexed page number>,
      "quote": "<exact verbatim substring from the parsed text>",
      "relevance": "<explanation of what this value/quote means and how it supports the answer>"
    }
  ]
}
ANSWER_EOF

Critical rules for the answer:

• Embed inline citation markers like [1], [2], etc. in your answer text, corresponding to the 1-indexed position in the citations array
• Place markers at the end of the sentence or claim they support
• Example: "Reserve Bank credit totaled **$6,613,609 million** [1], with securities held outright at $6,375,679 million [2]."

Critical rules for what to cite:

• Cite the EVIDENCE, not just the label. The user wants to audit your claims. If you say "revenue was $1.2B", cite the actual number 1,200,000 from the text — not just the heading "Revenue". You can cite both the value and the label if they're on the same page.
• Cite specific data values — numbers, percentages, dates, dollar amounts, quantities. These are what the user needs to verify.
• Each relevance field should explain the "so what" — not just restate the label but explain what this value means in context and how it supports your answer. E.g., instead of "Total revenue figure" write "Total revenue for Q3 2025, representing a 12% year-over-year increase that supports the growth trend discussed above."
• Include 5-15 citations covering all key claims and data points in your answer.

Critical rules for quote format:

• quote MUST be copied character-for-character from the parsed text. It is used for bounding box lookup via exact string matching. Do NOT paraphrase, reword, clean up, or fix typos.
• Prefer short, precise quotes — a single value like $2,769.23 or a short contiguous phrase like Securities held outright (under 40 characters). Shorter quotes match bounding boxes much more reliably.
• NEVER quote across table columns. In tabular data, each column cell is a separate text region. Quote individual cell values (e.g., $16,615.38), NOT a run of values spanning multiple columns (e.g., $2,769.23 80.00 $2,769.23 $16,615.38). Large whitespace gaps between values mean they are separate bounding boxes that won't match as one quote.
• If the text has unusual characters, hyphens, or formatting artifacts, include them exactly as they appear.
• page is 1-indexed (matches LiteParse pageNum)
• file is just the filename (not the full path)
• For plaintext files (.txt,.md), set page to 0 (they have no pages)

Step 4 — Generate HTML Report

Run the bundled script in generate mode to produce the visual report:

python "${CLAUDE_SKILL_DIR}/scripts/generate_report.py" \
    --skill-dir "${CLAUDE_SKILL_DIR}" \
    --dir "$0" \
    --answer-file /tmp/research_docs_answer.json \
    --output research_docs_output/

This will:

1. Parse and screenshot only the cited pages (efficient — not all pages)
2. Find bounding boxes for each cited quote
3. Generate a self-contained HTML report with the answer, page images, and highlights
4. Open the report in the default browser

Step 5 — Present Results

Tell the user:

1. Where the report was saved (the file path printed by the script)
2. A brief summary of the answer (2-3 sentences)
3. How many citations were found