The Protected Terms List: How to Keep Critical Details Intact During AI Rewrites

A protected terms list keeps names, numbers, quotations, and required language intact while AI helps revise the rest of a high-stakes draft.

A rewrite can improve a sentence and damage the information inside it at the same time.

A product name gains an extra space. A percentage becomes a rounded estimate. A direct quotation turns into a smoother paraphrase but keeps its quotation marks. A policy phrase is replaced with a friendly synonym that no longer carries the approved meaning. The new draft may read better while becoming less reliable.

A protected terms list prevents that quiet drift. Before revision begins, the editor records the details that must survive exactly, explains why they are fixed, and defines any changes that are allowed. AI can then help with structure, rhythm, and clarity without being asked to guess which strings are untouchable.

Rewriting Should Be Controlled Change

Most rewrite requests describe the desired movement: make this clearer, shorter, warmer, or more direct. They rarely describe the boundaries. That leaves the tool to treat every part of the draft as equally flexible.

For low-stakes copy, that may create only small annoyances. In technical, commercial, academic, policy, or regulated material, a small substitution can alter eligibility, attribution, price, scope, or responsibility. Fluency makes the problem harder to spot because the changed detail still sits inside a polished sentence.

A protected terms list changes the assignment. It says, in effect: improve the route, but do not move these landmarks. The list is not a promise that an automated rewrite will preserve everything perfectly. It is a clear specification that makes mistakes easier to prevent and much easier to detect.

Separate Fixed Content From Flexible Content

Begin by marking the draft in two layers.

Fixed content contains details whose exact form or exact meaning is controlled by evidence, identity, agreement, policy, or an approved source. It should change only through a deliberate decision by the appropriate reviewer.

Flexible content carries the explanation around those details. Sentence order, transitions, examples, rhythm, headings, and supporting language can usually move as long as the meaning remains accurate.

The distinction belongs to the specific document, not to a universal dictionary. The number “30” may be decorative in a brainstorm and critical in a cancellation policy. A phrase may be flexible in a blog post and mandatory in a customer notice. Editors need to ask why a detail is present and what would happen if it changed.

A useful first pass is to highlight names, identifiers, numbers, units, dates, quotations, citations, links, and formally approved phrases. Then inspect each one. Some will prove flexible after all. The rest become the protected list.

Build The List In Four Practical Buckets

A short structured inventory is easier to use than a warning such as “keep all important details.” Four buckets cover most editorial work.

1. Names And Identifiers

Protect official product names, organization names, plan names, model names, account types, document titles, reference codes, version numbers, file names, URLs, email addresses, and API paths when they must appear in an exact form.

AI rewrites often normalize capitalization, add or remove punctuation, expand abbreviations, or replace a formal name with a natural-sounding description. Those choices may be stylistically reasonable and operationally wrong. “Team Plus,” “Team+,” and “the team plan” are not interchangeable if only one is the real product name.

2. Figures, Dates, And Units

Protect prices, percentages, quantities, measurements, thresholds, time windows, effective dates, sample sizes, ranges, and version numbers. Include the unit and qualifier, not just the digits.

“Up to 10 MB,” “10 MB,” and “10 MiB” express different things. So do “within 30 days,” “after 30 days,” and “30 business days.” A rewrite that keeps the number while changing its condition has not preserved the fact.

The source should be recorded beside any important figure. Protection is not verification: an outdated number copied perfectly is still outdated.

3. Quotations And Citations

Direct quotations need exact wording, exact boundaries, and a checked source. Citations need the correct author, title, locator, link, and relationship to the claim they support.

Do not ask a rewriting tool to “improve” text inside quotation marks. If a quotation is too long or unclear, shorten it with proper omission practices, paraphrase it without quotation marks, or choose a better passage after returning to the source. Never preserve the appearance of a quote after its wording has been invented or polished.

4. Mandated Or Approved Terminology

Some language is controlled by a contract, policy owner, legal reviewer, standards body, subject-matter expert, accessibility guide, or brand system. Record that wording exactly and identify who can approve a change.

This bucket may include eligibility language, safety instructions, privacy terms, disclosure statements, defined technical terms, or the approved description of a feature. AI can clarify the surrounding explanation, but it should not silently substitute a near-synonym for language whose precision has already been reviewed.

Record Permitted Variants Instead Of Guessing

Not every protected item must appear in one rigid form everywhere. A full organization name may be required on first mention while an approved abbreviation is allowed later. A date may use a long form in body copy and an ISO format in metadata. A unit may have an accepted symbol.

Add permitted variants to the list explicitly:

• Official form: the exact wording used on first mention
• Allowed variants: approved short forms, symbols, or later references
• Not allowed: common but incorrect spellings, labels, or conversions
• Reason: identity, evidence, policy, contract, citation, or style rule
• Owner: the person or role authorized to approve a change

This prevents two opposite failures: freezing harmless variations that would improve readability, and allowing a rewrite to improvise where exactness matters.

Give The Rewrite A Bounded Instruction

Once the list exists, describe both the edit and its limits. A useful instruction names the audience, the desired improvement, the protected items, and what to do when a conflict appears.

For example: “Revise this product guide for a first-time customer. Shorten repeated setup, use direct sentences, and improve transitions. Preserve every protected string exactly, including capitalization, punctuation, numbers, units, URLs, and quotation marks. You may move a protected string to a clearer sentence, but do not paraphrase it. If clarity appears to require changing a protected item, flag the conflict instead of resolving it.”

Place the protected list beside the instruction, not in a distant document the tool cannot see. For a long draft, work section by section so the boundaries remain visible. Ask for a separate conflict list rather than allowing uncertain changes to disappear into revised prose.

Even a strong instruction is a control, not a guarantee. The output still needs comparison and editorial review.

Compare Exact Strings, Then Review Meaning

Use two checks because each catches a different class of error.

The first is a deterministic comparison. Search the revised draft for every required string and count its occurrences. Compare numbers, dates, units, names, identifiers, links, and quoted passages against the approved inventory. A script, spreadsheet check, document comparison, or careful search can do this. The method matters less than producing a clear pass-or-fail result for each protected item.

Exact comparison catches a missing hyphen, changed digit, altered path, or shortened quote. It does not prove that the surrounding statement is still true.

The second check is a human meaning review. Read the sentence around each protected item and ask:

• Does the same condition still apply to this figure or phrase?
• Did a transition change cause, timing, scope, or certainty?
• Does the citation still support the claim now attached to it?
• Is a direct quote represented fairly in its new context?
• Did the rewrite create a promise the approved wording never made?

A protected sentence can remain letter-for-letter identical and become misleading after the paragraph around it changes. That is why string preservation and meaning preservation must be separate gates.

Example: A Product Release Guide

Imagine a release guide that contains an official feature name, a minimum supported version, a file-size limit, an API path, and a quotation from the release notes. The original draft is accurate but repetitive, so an editor wants a clearer rewrite.

The protected list stores the feature name with its approved capitalization, the complete version number, the limit with its unit and qualifier, the exact API path, and the verified quotation with its source link. The editor permits an approved short feature name after first mention but does not permit rounding the version or describing the limit as a recommendation.

After the rewrite, an exact comparison confirms that every protected string remains. The meaning review then catches a subtler problem: the new introduction implies that the feature is available to every account, although the preserved release-note quote refers only to eligible accounts. The prose is revised again to restore the missing condition.

The list protected the pieces. Human review protected the conclusion.

Example: An Internal Policy Summary

Now imagine a team turning an approved policy into a shorter employee summary. The source includes an effective date, a response window, a defined role, and one paragraph that the policy owner requires the summary to reproduce exactly.

Those details go on the protected list with the policy owner named as the change authority. Headings, explanations, and examples remain flexible. If the rewrite makes the mandated paragraph feel abrupt, the editor improves the lead-in instead of editing controlled language without approval.

The final review compares the summary with the governing policy, not merely with the previous draft. That distinction matters: a protected list prevents accidental mutation during revision, but it cannot certify that the starting text was current, complete, or appropriate for the audience.

Do Not Lock What Has Not Earned Protection

An overgrown list can make a rewrite useless. If every sentence is protected, the editor has created an archive, not a revision brief.

Do not lock filler because it survived several drafts. Do not preserve a number before checking its source. Do not freeze jargon merely because it sounds technical. Do not treat AI-generated wording as authoritative because it arrived with confidence. And do not use the protected list to avoid a necessary review by legal, compliance, policy, or subject-matter owners.

Protection should follow a reason. If no one can explain the consequence of changing an item, it may belong in the flexible layer. If the consequence is real, record the source and owner so future editors understand the boundary instead of inheriting unexplained rigidity.

The Final Protected Terms Checklist

Before approving an AI-assisted rewrite, confirm:

• Fixed and flexible content were separated before rewriting
• Names and identifiers match their authoritative source
• Figures retain their dates, units, conditions, and qualifiers
• Quotations are exact and citations still support the attached claims
• Mandated language changed only with the correct approval
• Permitted variants and prohibited substitutions are explicit
• The rewrite instruction says what to preserve and how to flag conflicts
• An exact comparison passed for every required string
• A human reviewer checked meaning, scope, and context
• The protected list records a source, reason, and owner where needed

A responsible rewrite is not one that changes as much as possible. It changes the parts that benefit from revision while keeping verified details under deliberate control.

The protected terms list makes that control visible. It gives AI room to help with language, gives reviewers a concrete quality check, and keeps precision from becoming the hidden price of smoother prose.