The legal-document engine

LegalWork's actual legal work is defined as firm-owned agent prompts, not app code. This path walks the two flagship workflows — Word redlining and tabular review — from the user command down to the read-only subagents that do the work.

3. Fan out — one subagent per document, in parallel

For each document, spawn the document-extractor subagent with the Task tool. Issue the calls in parallel (multiple Task calls in a single turn) so the grid fills concurrently — this is the whole point of the fan-out.

Default granularity is one extractor per file: it reads the document once and fills that file's entire row across all columns. (This is cheaper and more consistent than one subagent per cell, because the document is read a single time.) Escalate to a per-cell extractor only for a column that is high-stakes and came back low confidence or conflicted — re-run just that (file, column) pair with a focused prompt.

Give each extractor exactly this:

FILE: <path to the one document>
DOC_TYPE: <NDA | Commercial Lease | ... | unknown>
COLUMNS:
  1. key: parties
     question: Who are the parties to this agreement (full legal names)?
     hint: Check the preamble and signature block.
  2. key: governing_law
     question: Which jurisdiction's law governs?
  ... (every column, with key + question, hints where useful)

Return ONLY the strict JSON object defined in your instructions.

Each extractor returns a JSON object: { file, title, docType, summary, cells: [ {key, value, reason, quote, page, location, confidence} ] }. Remember: value is short (it's a table cell), reason is the longer sidebar explanation, quote is one verbatim sentence, and page is the 1-based page that sentence is on.