# Large File Review Playbook

Version 1.0 · Template for local adoption · Suggested owner — data or documents editor

## Intake and baseline

Record the question, desired deliverable, exact source versions and what “entire file” includes. Default to all pages or structural blocks, all sheets including hidden sheets, and all records including filtered or hidden rows. For workbooks, inventory all non-empty cells in each sheet, formulas, and any other material required for the assignment. Define a bounded rectangular range when empty cells matter. Enumerate workbook objects such as comments, charts and images when they may carry evidence and document any limitation inspecting them. Record explicit exclusions. Keep originals intact. Use original filenames and stable file IDs. Record hashes when feasible and mark when a chunk or export cannot be matched to the original.

Use a reproducible parser or reader for the original file, or an approved local extraction plus an independent manifest. Establish expected totals from the original file metadata, a separate enumeration or a verified human manifest. Do not estimate from samples, visible scroll limits, retrieval hits or previews. Compare observed processing totals with that baseline. If the baseline is unavailable, mark expected counts unknown. For file constraints, check the current platform documentation and test with a small representative chunk. [Working with files in ChatGPT](https://learn.chatgpt.com/docs/artifacts-viewer) · [Spreadsheet analysis example](https://learn.chatgpt.com/use-cases/clean-messy-data)

## Preserve structure by format

- **PDF.** Count viewer pages, including blanks, covers and appendices. Read each page in order and assign original viewer page numbers. Keep printed page labels separately. For scan-heavy or visual pages, use approved OCR and/or page images, preserve the page mapping and check the rendered originals for consequential details. When the assignment depends on page images or embedded charts, test a representative page on the intended account and upload route. If the file is available as text only, use approved page-image attachments or OCR with an original-page manifest and inspect the visuals separately. Process dense material in ordered, verifiable chunks. Check visual-only charts, figures and table content separately. A blank page, an image-only page and a failed extraction require distinct statuses.
- **DOCX.** Enumerate paragraphs, headings and tables in document order, including nested tables and in-scope notes, headers, footers, text boxes, embedded objects or tracked revisions. Record the agreed revision view and any object types the parser cannot see. Use structural locators. If a fixed PDF rendition supplies page anchors, identify its version and how its pages relate to the original. Inspect visual elements using an approved rendition where needed.
- **TXT.** Confirm encoding and line endings. Preserve original line numbers and account for blank lines. Record decoding errors.
- **CSV.** Confirm delimiter, encoding, quoting, header and field count. Parse logical records, including quoted multiline values. Preserve physical line spans if available. Keep original order, raw strings, leading zeros, duplicates and blank fields. Count parsed, empty and malformed records separately. Record schema deviations without silently repairing or dropping them.
- **XLSX.** Enumerate sheets in workbook order and inventory their visibility and in-scope ranges. Use original coordinates, preserve hidden and filtered rows, merged regions, types, leading zeros and distinctions among empty cells, errors, formulas and displayed or cached values. Record the chosen value basis and any missing or stale formula cache. For calculated claims, document any explicitly requested, reproducible recomputation. Avoid executing macros or following external links. If a workbook is converted, use a manifest and preserve sheet names and original coordinates. Record which formulas, comments, visuals or other objects the export omitted.

## Phase 1 · Build and reconcile the contents map

1. **Plan ordered chunks.** Use consecutive original units sized for the tool. Give each a chunk ID, expected range and counts. If a split needs repeated headers or overlap, mark them as context and count original units once. For spreadsheets, enumerate the full bounded range or stream all records and retain original row and cell coordinates.
2. **Process each chunk in order.** Inspect every in-scope unit using a method appropriate to the assignment. For structured files, use ChatGPT data analysis if it can read the original file, or use an approved local parser. Read every original unit in the bounded scope, preserving values and exceptions. Save the method, counts and output so the editor can reproduce the result. For a semantic or visual task, inspect the actual ordered units in reviewable chunks. Parsing or listing file contents establishes enumeration only. Record a short section or content description, an evidence marker or an explicit no-match status for each unit, at a granularity that allows the editor to check coverage. Keep the ledger as a separate CSV or workbook when supported.
3. **Build the contents map before delivering analysis.** For narrative material, show headings or descriptive labels, start and end locators and annexes. Mark headings added by the reviewer. For data, show sheet or section order, headers, original ranges, row/field or cell counts and visibility. For multiple files, begin with the manifest and include a map per file.
4. **Reconcile.** Compare original expected units to unique reviewed units, enumerated-only units, failed units and agreed exclusions. Check for gaps, overlaps, duplicates, off-by-one errors and version mismatches. Use set or range comparisons when code is available. Recheck first and last units, boundaries, errors and samples against the originals. A sample is a quality check alongside the full pass.
5. **Checkpoint and resume.** Save the manifest, map, ledger, processing method or script and any intermediate evidence tables to the authorized working location or return them for the reporter to save. Resume using the same versions and validated IDs. State last completed and next expected ranges. If interrupted, output “Partial review” with exact missing or unverified ranges. Withhold full-coverage claims until the gaps are resolved. If a complete pass cannot be made in ChatGPT, request smaller ordered chunks, missing files, a better scan or an approved local parser and manifest. Report which work remains.

### Templates

**Manifest.** File ID | name and version | original identifier or hash | type | expected unit totals and baseline method | scope and exclusions | chunk plan.

**Contents map.** File ID | section/sheet | original start–end | label or headers | expected and reviewed counts | exceptions.

**Coverage ledger.** File ID | chunk | original unit or range | expected and observed counts | enumeration/review method | status | evidence marker or outcome | exceptions/next step. Use `Reviewed`, `Reviewed — no match`, `Confirmed blank`, `Enumerated only`, `Unreadable`, `Missing` or `Excluded by scope`. Keep uncertain OCR or material conversion losses explicit. Only reviewed or confirmed blank units count as complete within scope.

**Extraction record.** Finding ID | extracted value or short verbatim excerpt | exact source file and original locator | qualifying context | method or formula if derived | confidence limitation.

## Phase 2 · Extract or analyze

After the contents map and coverage ledger are ready, apply the reporting question to the reviewed source material. Read the original units or a versioned, traceable extraction of them. Use the contents map to navigate and the underlying records to verify findings. Apply the documented selection rule to the full applicable unit set for an exhaustive extraction, even when the map suggests that matches are concentrated in one section. For semantic extraction, log chunks checked for the specific question and record no-match outcomes. Keep extraction completeness separate from the completeness of the initial map. If the first phase is partial, continue with a scoped result only where the missing units do not undermine the requested conclusion. Show any remaining gaps prominently.

## Deliverable order

1. Source manifest and table of contents or workbook map.
2. Coverage totals and exceptions, plus the detailed ledger or a reference to it. Name the tool or method actually used and distinguish enumeration from review.
3. Requested extraction or analysis. Provide an original anchor and brief evidence for every extracted item. For aggregates, attach the source rows or a reproducible range and selection rule, and include units, denominator, handling of missing values and calculation.
4. Gaps, next chunks and editor checks. Tie absence findings to the verified scope.

## Copy-ready journalist prompts

> Review all of [file]. First build its contents map and reconcile a page/block ledger against the original. Then extract [target] with original locators and short excerpts. Account for appendices, blanks, scans and visuals. Work only with this file. If any part is inaccessible, name the exact ranges and mark the result partial.

> Review [workbook] in full, including hidden sheets and rows. Build a workbook map and reconcile sheet, row and in-scope cell totals. Extract [fields or criteria] with original sheet/cell references and evidence values. Preserve raw IDs and explain formula and missing-value handling.

> Resume the review using [manifest, checkpoint, and source chunks]. Verify versions, original ranges and the last completed unit. Process the missing units in order, reconcile the ledger and produce the requested analysis only for the scope you can support.