tabular-review

/tabular-review

1. Load ~/.claude/plugins/config/claude-for-legal/corporate-legal/CLAUDE.md → diligence structure, thresholds, house format.
2. Confirm: what documents, what columns, where does the output go.
3. Build the typed schema. Write .review-schema.yaml. Confirm with the user.
4. Sample run (3–5 docs). Adjust schema. Confirm.
5. Fan out — one sub-agent per document, parallel. Each cell: value + state + verbatim quote + location.
6. Normalization pass. Flag outliers and inconsistencies.
7. Output: .xlsx or Google Sheets (ask which), plus .csv + _sources.csv + markdown always. Work-product header.
8. Summary: verification workload (counts of not_present / unclear / needs_review per column), flagged columns, where the files are, reminder that every cell is a lead not a finding.

/corporate-legal:tabular-review
/corporate-legal:tabular-review --schema .review-schema.yaml --docs ./vdr/02-Contracts/
/corporate-legal:tabular-review --template ma-diligence

--schema <path>: Use an existing schema file instead of building one. Useful for re-runs and incremental additions.

--template <name>: Start from a template in references/. Currently: ma-diligence.

--docs <path>: Document source. A local folder, a Drive folder ID, or a VDR path. If omitted, asks.

--output <xlsx|gsheets|csv>: Output format. If omitted, asks.

--sample <n>: Sample size for the schema check. Default 5.

Matter context

Matter context. Check ## Matter workspaces in the practice-level CLAUDE.md. If Enabled is ✗ (the default for in-house users), skip the rest of this paragraph — skills use practice-level context and the matter machinery is invisible. If enabled and there is no active matter, ask: "Which matter is this for? Run /corporate-legal:matter-workspace switch <slug> or say practice-level." Load the active matter's matter.md for matter-specific context and overrides. Write outputs to the matter folder at ~/.claude/plugins/config/claude-for-legal/corporate-legal/matters/<matter-slug>/. Never read another matter's files unless Cross-matter context is on.

Purpose

You have a pile of documents and a list of questions you need answered consistently across every one. A diligence request list. A vendor contract audit. A lease portfolio review. The output is a table: document rows, data-point columns, and every cell traceable to the exact words in the source.

This is not issue spotting. diligence-issue-extraction finds the 30 problems hiding in 2,000 documents. This skill answers the same 15 questions about all 2,000 documents. Both are legitimate; they answer different questions.

This is also not a replacement for a human reading the document. Every cell this skill produces is a lead that needs verification, not a finding. The output is designed to make verification fast, not to skip it.

Load context

• ~/.claude/plugins/config/claude-for-legal/corporate-legal/CLAUDE.md → diligence structure, materiality thresholds, house format preferences
• ~/.claude/plugins/config/claude-for-legal/corporate-legal/deals/[code]/deal-context.md if working a specific deal
• An existing schema file if the user has one (.review-schema.yaml)

The column type system

The thing that makes a tabular review useful is that Column C means the same thing in row 1 as in row 200. Free text drifts. Types hold.

Every column has a type that constrains the answer format:

The verbatim rule: Every non-verbatim column also captures the exact source quote that supports the answer, as a companion field. The answer in the cell is the interpretation; the quote is the evidence. A classify cell that says "consent not unreasonably withheld" is useless without the sentence it came from, because the reviewer's job is to check whether that's the right read.

The three states of "not found"

A blank cell hides information. Force one of three explicit states whenever you can't produce a positive answer:

These are three different pieces of information. A deal team handles "the contract is silent on assignment" very differently from "the assignment clause is ambiguous." Collapsing them into one blank cell loses the distinction.

Workflow

Step 0: What and where

Confirm:

1. Documents. Where are they? VDR MCP (Box, Datasite, iManage), local folder, Google Drive folder, or a list of files. How many? If >200, warn that this will take a while and offer to start with a materiality-filtered subset.
2. Schema. What columns? Two paths:
   • User picks a template from references/ (M&A diligence standard is the default)
   • User describes columns in natural language and you structure them into the typed schema
3. Output. Excel (.xlsx) or Google Sheets — ask which the team works in. CSV and markdown always written as fallbacks. Output goes to the deal folder, Drive, or wherever the user says.

Step 1: Build and confirm the schema

Turn the user's column list into a structured schema. For each column: a stable id, a human label, a type, a prompt (the question a reviewer reading the document would ask), and for classify columns an options list.

Write it to .review-schema.yaml next to the output. This file is the reusable artifact — the user can edit it, add a column, re-run against new documents. Show it to the user and confirm before fanning out.

schema:

Step 2: Sample run

Do not fan out to 200 documents on an untested schema. Run 3–5 documents first. Show the user the rows. Look for:

• Columns where most answers are unclear — the prompt is ambiguous, rewrite it
• classify columns where answers don't fit the options — add options or change to free
• verbatim columns returning paraphrases — reinforce that it must be character-for-character

Adjust the schema, re-run the sample, confirm. This saves the user from a full run that has to be thrown out.

Step 3: Fan out

One sub-agent per document, in parallel. Each sub-agent:

1. Reads the entire document (not a RAG chunk — the whole thing).
2. For each column, finds the relevant provision.
3. Returns a structured row: for each column, {value, state, quote, location}.
   • value is the typed answer (or null if state is not answered)
   • state is answered | not_present | unclear | needs_review
   • quote is the verbatim supporting text (exact, no paraphrase, no ellipsis inside a sentence — if you cut, cut at sentence boundaries and mark it)
   • location is where the quote lives (section number, heading, page — whatever the document gives you)

The quote is not optional, and the verbatim rule is mechanical, not exhortation. Each sub-agent must comply with all of the following before returning a cell with state: answered:

• The quote MUST be a character-for-character copy of contiguous text from the source document, retrievable at the location the sub-agent cites. Do NOT compose a quote from a section heading plus standard boilerplate you expect to be there. Do NOT paraphrase and call it verbatim. Do NOT reconstruct a quote from memory of how such clauses "usually" read. Do NOT fill gaps in the source with ellipsis-stitching across non-contiguous text.
• The location must be specific enough for the normalization pass to re-open the document and re-read the same span — a section number, heading, or page reference the reviewer can navigate to.
• If the sub-agent cannot locate and copy the exact text (source truncated, OCR garbage, provision implied but not written, section heading visible but body not loaded), the cell state is needs_review, the value is null, and notes MUST contain quote_unavailable: <reason>. It is NEVER acceptable to set state: answered with a composed or reconstructed quote.
• The same rule applies to verbatim-typed columns AND to the companion source quotes attached to classify / date / duration / currency / number / free cells. The supporting quote carries the same verbatim obligation as the cell value.

The normalization pass in Step 4 spot-checks this by re-reading the source at the cited location and comparing the stored quote character-for-character against the source text. A mismatch downgrades the cell to needs_review, notes quote_mismatch, and flags the whole column for a wider spot-check — if one sub-agent composed a quote, others in the same run may have too.

Step 4: Normalize

After the fan-out, read the whole table column by column. This is the pass that catches the failure mode of every tabular review tool: the same clause interpreted inconsistently across documents.

For each classify column:

• Check that every answered value is in the options list. Outliers get re-classified or bumped to needs_review.
• Check for clusters: if 180 documents say consent_required and 20 say consent_not_unreasonably_withheld, that's probably real. If 195 say consent_required and 5 say freely_assignable, look at the 5 — they're either genuinely different or misclassified.

For each date / duration / currency column:

• Check format consistency. Normalize.
• Flag implausible values (a 99-year term, a $1 cap) as needs_review.

For each verbatim column AND for the companion source quotes on every other column:

• Spot-check by re-opening the source document at the cited location for a random sample (at least 3–5 rows per column, or 10% of rows, whichever is larger) and comparing the stored quote character-for-character against the source.
• If any quote is composed, paraphrased, reconstructed, or cannot be located at the cited span: downgrade that cell to needs_review with quote_mismatch in notes, and flag the whole column — expand the spot-check to the rest of the column rather than assuming the other rows are clean. One fabricated quote is enough to justify widening the check.
• A cell with state: answered and a mismatched quote is a higher-severity failure than an unclear or needs_review cell — it misrepresents the evidence trail. Downgrade aggressively.

Step 5: Output

Write the table in three formats:

Markdown (always, for in-session review):

CSV (.csv, always): One file for the values, one companion file for the quotes and locations (_sources.csv). Keeps the main file clean and the evidence trail complete.

Excel (.xlsx) or Google Sheets — whichever the user works in. Ask; don't guess. Both follow the same workbook structure (see references/excel-output.md and references/gsheets-output.md). For Excel: Claude in Excel (Office agent) if available, openpyxl fallback. For Sheets: Sheets MCP if available, Sheets API via ADC, CSV-import fallback. In the spreadsheet output:

• Each data column is paired with a hidden source column containing the quote and location. Cell comments (Excel) or notes (Sheets) on the visible column surface the quote on hover.
• Color code by state: white = answered, yellow = unclear or needs_review, gray = not_present.
• A Verified column per data column, blank by default. The reviewer marks it. This is the verify/flag pattern that makes the table auditable — the deal team can see at a glance what a human has actually checked.
• A _schema sheet with the column definitions, so the file is self-documenting.

Prepend the work-product header from the plugin config ## Outputs as a top row. Alongside it, include a distribution note:

This review is derived from source documents that may be privileged, confidential, or both. It inherits the sources' privilege and confidentiality status — distribution beyond the privilege circle can waive privilege. Store with the matter's privileged files and make distribution decisions deliberately.

Step 6: Summary

After the table is written, give the user a one-screen readout:

• Document count, column count, rows completed
• Count of not_present, unclear, needs_review per column — this is the verification workload
• Any columns where the normalization pass flagged >10% of rows
• Where the output files are
• A reminder: every cell is a lead, not a finding. Verification required before this informs a rep, a schedule, or a memo.

Close with the next-steps decision tree

End with the next-steps decision tree per CLAUDE.md ## Outputs. Customize the options to what this skill just produced — the five default branches (draft the X, escalate, get more facts, watch and wait, something else) are a starting point, not a lock-in. The tree is the output; the lawyer picks.

What this skill does not do

• It does not replace reading the documents. It tells you where to look.
• It does not produce confidence scores. A 0.73 is not information. The unclear / needs_review states and the verbatim quotes are the confidence signal — if the quote doesn't support the value, flag it.
• It does not silently skip documents. Every document the user pointed at gets a row. A document that couldn't be read gets a row of needs_review with a note.
• It does not pretend a paraphrase is a quote. The evidence trail is the whole point.

Relationship to other skills

• diligence-issue-extraction finds issues; this extracts data points. If an extraction reveals an issue (a MAC clause that references a specific earnings target, a poison pill), note it and suggest running diligence-issue-extraction on that document.
• material-contract-schedule builds one specific table (the disclosure schedule). It can consume this skill's output directly — the schedule is a filtered, reformatted view of a tabular review.
• ai-tool-handoff hands bulk review to Luminance/Kira when the corpus is too large or the team prefers a dedicated platform. This skill is the in-house option for anything it can handle — run it first, hand off the residue.

Output safeguards

Every output gets the work-product header. Every cell gets a source citation or a flagged state. The summary explicitly says verification is required. The Excel Verified column makes the verification state auditable. This is not a tool that lets you skip reading; it's a tool that makes reading faster.