GPT 6 document formatting with preservation checks

Define the GPT 6 document formatting boundary

GPT 6 document formatting is illustrated in OpenAI’s Astra launch material through a legal-document formatting demonstration and a document-styling example. OpenAI also describes improved template adherence. This establishes the official example context without showing that every document, style guide, or export format will be preserved correctly. This workflow is original guidance and has not been run against a live GPT-6 system. It assumes that the source document and output are available in an approved tool. The goal is to apply a style contract while preserving content and making every ambiguous choice visible.

List what must remain unchanged: wording, headings, order, tables, numbers, links, images, captions, footnotes, and metadata. Then list what may change: typeface, spacing, margins, colors, headers, footers, and page breaks. If the brief does not answer a question, record it as unknown.

Write a GPT 6 document formatting prompt

Apply this style guide [STYLE GUIDE] to [DOCUMENT]. Preserve meaning, wording, headings, tables, images, links, captions, footnotes, and required order. Return a change ledger, an ambiguity list, and a validation checklist. Do not rewrite content or invent missing captions. If a style rule conflicts with content preservation, stop and ask for a decision.

Ask for a structural inventory before transformation. The inventory should count headings, tables, images, links, notes, and sections. Compare that inventory with the output. A count mismatch is a review signal, not an invitation to guess what was lost.

Illustrative worked example

This example is fictional and demonstrates the acceptance shape only. Input: a three-page brief with four headings, one two-column table, two links, and one captioned image; style guide: use a 12-point body, 18-point section headings, 1-inch margins, and blue links. The expected output keeps four headings, one table, two working links, and one image with its caption while applying the styles. Acceptance includes opening the exported file, checking a dense table page, activating both links, confirming reading order, and recording any page-break exception. A model may draft the ledger; an editor verifies the actual file.

Visual review

Open the output in the target application and inspect the first page, a dense page, a page with a table, a page with an image, and the final page. Check margins, line wrapping, page breaks, headers, footers, captions, list indentation, and orphaned headings. Export to the required format and inspect that version too.

Visual review should include long headings, long URLs, special characters, missing fonts, and narrow screens when relevant. A preview can hide a broken link or an accessibility order problem. Verify links and document navigation separately.

Failure modes

Common failures include applying a heading style to ordinary text, changing a number while reflowing a table, dropping an image anchor, flattening a list, or replacing a link with visible text. A model may also “improve” prose even when the task is formatting. Keep content edits outside the formatting pass unless the owner explicitly approves them.

If a table does not fit, record the options: landscape section, smaller type, split table, or an approved layout change. Do not silently remove columns. If a page break produces a confusing reading order, flag it for a person to resolve.

The GPT 6 document formatting acceptance record

Record source filename and version, output filename and version, style rules applied, preserved object counts, visual pages inspected, links checked, accessibility checks, and unresolved questions. Have the content owner approve any change that could affect interpretation. Keep the original file unchanged and retain the output as a separate version until approval.

The document reconciliation workflow helps when two formatted versions also contain content changes. The GPT 6 guide explains how to keep generated drafts and human decisions separate.

A useful stopping rule

Stop when the output satisfies the style contract, the preservation inventory matches, required visual samples pass, links and navigation work, and the owner accepts the remaining ambiguities. Formatting is complete when the file is usable in its target context, not when a preview merely looks tidy.

Preserve structure during GPT 6 document formatting

Use real heading levels, list structures, table headers, captions, and link destinations instead of simulating them with bold text or spacing. Check reading order in the target application and, when required, with an accessibility tool. Confirm that color is not the only way to distinguish status. A visual style guide describes appearance; a usable document also needs semantic structure.

Watch for transformations that change the source unintentionally. Copy and paste can alter smart quotes, minus signs, decimal separators, or hidden characters. Reflow can move a footnote away from its reference. Export can rasterize text or remove a link. Compare important fields and search for required terms after the transformation.

Treat the change ledger as part of the deliverable. Say which rules were applied, which items were skipped, and which decisions remain open. A later editor can then update formatting without repeating the investigation. When the owner approves, archive source, output, ledger, and validation record under the project retention policy.

Review representative pages

Choose pages that exercise different structures rather than reviewing only the opening page. Include a heading dense page, a page with a long table, a page with an image or caption, a page with lists and links, and the final page. Check the output at the intended print or screen size. If a document has sections with different headers, margins, or numbering, inspect the transition between them.

Keep a preservation count for headings, tables, images, links, notes, and pages. Compare it after export and investigate every difference. For documents that will be shared broadly, also check title, language, reading order, and metadata. These checks help a person catch a technically valid transformation that is still confusing or inaccessible. For release handoff, include the source and output versions, style guide, preservation counts, representative pages, export format, and owner approval. If the file will be printed or passed through another editor, test that path before calling it ready. A clear record makes a later correction local and recoverable, rather than forcing the team to guess which transformation produced the current layout. Keep the prompt, source inventory, and change ledger together. When a rule is added, state which objects it affects. When an exception is approved, record the owner and reason. When an export path changes, rerun the representative page review. This gives the next editor a reliable starting point and limits accidental changes to meaning. Finally, ask whether the formatted file serves its real reader and delivery path. If visual polish makes the document harder to navigate, change the style decision. If preservation checks pass but the page is still confusing, record a content or information design issue separately. This preserves the formatting scope while giving the owner a useful next decision. The point of formatting is a usable, approved file. Preserve the original and keep the transformation record with the output. A future editor should be able to identify what changed, what was preserved, which exceptions were approved, and which export path was checked. Before delivery, ask another editor to open the output from the handoff record alone. If they cannot identify the source, exceptions, or checked export, the record is incomplete. Improve it before distributing the file. That discipline makes the file maintainable. A later editor can repeat the visual checks, understand approved exceptions, and avoid changing substantive content while applying a new style. Use the same labels in the change ledger and handoff ticket so “proposed,” “observed,” and “verified” retain one meaning across the team. That shared vocabulary prevents visual confidence from being mistaken for a completed preservation or accessibility check.

A concrete GPT 6 document formatting run

Input: source filename and version, style guide, required output format, and preservation list. Ask GPT 6 document formatting to inventory headings, tables, images, links, captions, footnotes, and metadata before making changes. Return proposed style actions and any conflict between a style rule and content preservation.

Output review: compare object counts, search required terms, open the exported file in the target application, inspect representative pages, test links, and check reading order. The GPT 6 document formatting result is accepted only when visual and structural checks agree. If a table no longer fits, record the approved choice among landscape, split, or smaller type; never silently drop a column.

If an export changes minus signs, decimal separators, anchors, or footnote order, preserve the original and flag the output. If the style guide is ambiguous, return a decision request with the affected objects. This makes GPT 6 document formatting a reversible transformation with a clear acceptance record.