tabular-review

Tabular Review

This skill is how this firm runs tabular document review — the workflow Harvey and Legora call a "review grid": documents are rows, the fields you care about are columns, and every cell is an independent, source-cited extraction. Unlike the SaaS versions, this runs on the firm's own models and infrastructure, the column logic lives in firm-owned doctype skills, and the output is a self-contained artifact the firm keeps.

You are the orchestrator. You do not read the documents yourself. You define the grid, fan out one document-extractor subagent per document, then assemble the results into one HTML artifact.

The shape of the job

                 Parties     Term      Governing law   Assignment ...   ← columns (fields)
  acme_nda.pdf   [cell]      [cell]    [cell]          [cell]
  beta_msa.pdf   [cell]      [cell]    [cell]          [cell]      ← rows (documents)
  ...

Each cell = a short value (what shows in the grid) plus, behind a click, a longer reason, one verbatim quote sentence, and the cited PDF page rendered with that sentence highlighted. Every value is grounded in a quote from that document or it is "Not found". No hallucinated cells.

---

Workflow

1. Resolve the document set (rows)

Find the files to review. They may be attached, referenced by @path, named in the prompt, or sitting in a folder ("review the NDAs in ./ndas"). Use glob/list to expand folders. Confirm the list with the user if it's ambiguous or large (>~20).

Sniff the document type of each file (from filename and, if cheap, a first-page peek). You'll use this both to pick columns and to tell each extractor what it's looking at. A set can be mixed (some NDAs, some leases) — that's fine; group by type.

2. Resolve the columns (fields) — THIS IS THE BRANCH POINT

Columns can come from three places, in priority order:

1. The user already specified them. ("Extract party names, term, and governing law.") Use those verbatim; only add a column if you ask first.
2. A loaded doctype skill. Look for a skill named doctype-<type> (e.g. doctype-nda, doctype-commercial-lease). If one matches the documents, load it with the skill tool and use its recommended columns as the default set. List available skills first if unsure what exists.
3. Neither → ASK THE USER. Do not invent a column set silently. Detect the doc types, then ask what to extract and propose a starter set based on those types, using the suggestion library below. Make it a one-tap decision: offer the suggested columns and let them add/remove. Example: