AI edits create a perfect final draft

Achieving maximum clarity in your brief by editing for brevity and simplicity with AI tools

2026 May

Readers need clarity in your brief

Lawyers should aim to file a perfectly clear brief because judges – like all readers – need the writer’s help. As E.B. White wrote in The Elements of Style: “The reader is in serious trouble most of the time, floundering in a swamp, and it is the duty of anyone attempting to write English to drain this swamp quickly and get the reader up on dry ground, or at least to throw a rope.” Similarly, Jacques Barzun wrote in Simple and Direct that “the reader’s part... must never become a strain.” As former California Supreme Court Justice Marcus Kaufman told one of your authors: “Don’t write in a way that makes readers think, because they might not.”

Accordingly, Jonathan Franzen urges that “difficulty” for the reader is “a sign of trouble.” But you do not want your briefs to cause “difficulty” for the judge because the judge is already burdened by multiple difficulties – the judge may be unfamiliar with your area of law; the judge will be frequently interrupted; and the judge’s job – applying disputed legal principles to disputed facts – is inherently difficult. 

Moreover, judges work under exquisite time pressure. As Ruth Bader Ginsburg wrote: “Readers of legal writing...often work under the pressure of a relentless clock. They may lack the time to ferret out bright ideas buried in complex sentences, overlong paragraphs, or too many pages. Strong arguments can escape attention when embedded in dense or Delphic prose.” Lawyers should take care that unclear writing does not add to the judge’s burdens.

In this light, lawyers should write with maximum clarity to achieve maximum success. As Ruth Bader Ginsberg wrote: “Lawyers serve their clients best when their readers can quickly and firmly grasp their point... Lucid, well-ordered writing can contribute immeasurably to a lawyer’s success.” Judge Murry Cohen of the Texas Court of Appeals put Judge Ginsberg’s point more bluntly: “Good writing is rewarded so automatically that you don’t even think about it.”

The link between a clear writing style and success in court is recognized by famous writers and judges. Hemingway said: “The indispensable characteristic of a good writer is . . . lucidity.” C.S. Lewis wrote: “Take great pains to be clear. … [T]hough you start by knowing what you mean, the reader doesn’t.” Chief Justice Roberts explained that your “writing conveys not only your argument, but it also conveys a sense of your credibility and the care with which you put together your case.” With every brief you file your reputation is at stake, wrote Thoreau: “The best you can write will be the best you are . . . . The author’s character is read from title-page to end.”

Fortunately for lawyers seeking to write clearly, excellent books by respected judges and writers abound. B. Garner, Legal Writing in Plain English (2001); R. Flesch: How to Write Plain English: A Book for Lawyers and Consumers (1979); A. Scalia and B. Garner, Making your Case (2008); Securities and Exchange Commission: A Plain English Handbook: How to Create Clear SEC Disclosure Documents (1998); R. Wydick, Plain English for  Lawyers (2005); W. Zinsser, On Writing Well (2001).

Brevity

The authors of these texts agree that the prime ingredient of clarity is brevity. The most famous rule in The Elements of Style is “Omit needless words.” William Zinsser wrote that “clutter... blunts the painful edge of Truth... Writing improves in direct ratio to the number of things we can keep out of it.” Kurt Vonnegut urged writers not to flinch: “Have the guts to cut,” he urged.

Evidence that drastic cuts can improve the final draft is supplied by Robert Caro’s book, “The Power Broker,” a biography of Robert Moses, New York City’s public works czar. Caro’s “final” draft contained approximately 1 million words. But Caro and his editor deleted 350,000 words. As a result, the book won many prizes, including the Pulitzer Prize,

To achieve brevity, lawyers should delete: 1. Unnecessary words – e.g., very; clearly – that fail to advance your argument. 2. Redundant words. 3. Implicit words, telling readers what is apparent from the context. George Orwell advised writers to “never use a long word where a short word will do.” And Churchill urged that “short words are best and the old words when short are best of all.”

Simplicity 

The second ingredient of clarity is simplicity. Compare the readability of the following two jury instructions:

“Failure of recollection is common. Innocent misrecollection is not uncommon.” (BAJI 2.21; note double negative)

“People often forget things or make mistakes in what they remember.” (CACI 107). 

To achieve simplicity, imitate Justice Oliver Wendell Holmes: “The life of the law has not been logic; it has been experience.” Scalia and Garner warn, “never use a word that the judge may have to look up.” George Orwell warned: “Never use a foreign phrase, a scientific word, or a jargon word if you can think of an everyday English equivalent.” These warnings apply also to uncommon acronyms that the judge will not immediately recognize.

Editing: A hard task where AI can help

Lawyers can benefit from using editing assistance from AI because the lawyer – facing an impending deadline –will likely be deterred from careful editing by the exhausting effort already expended to research the law and facts and to write and revise the earlier drafts.

        Moreover, diligent editing is hard work. Hemingway said that to “get[] the words right,” he rewrote the last page of A Farewell to Arms 39 times. Isaac Babel explained: “I work like a pack mule,... I’m like a galley slave . . . . I go over each sentence, time and again.” Robert Caro had a mania for editing: “I rewrite the galleys completely. I even rewrite in page proofs, which they don’t allow you to do. I’d rewrite in the finished book if I could.”

And the final edit is the last chance to revise for clarity. As William Zinsser explains: “Rewriting is . . . when the game is won or lost. . . . The newly hatched sentence almost always has something wrong with it...[C]lear writing is the result of a lot of tinkering.”

In 1960 Sen. John F. Kennedy revised his famous ”ask not” challenge seven times before delivering at his inauguration this memorable exhortation:“[A]sk not what your country will do for you – ask what you can do for your country.”

The lawyers’ task is to instruct the AI to achieve brevity and simplicity in a style the judge will find acceptable. For example, younger judges will likely feel comfortable with such informal constructions as split infinitives (approved by the Chicago Manual of Style), ending sentences with prepositions (approved by E.B. White), and simple contractions (approved by Bryan Garner, but not by Justice Scalia). Older judges will likely find informal constructions offensive as lacking respect for the court. The solution is to know your audience – read your judges’ prior opinions and conform your writing to their preferences.

How AI works with prompts

AI writing tools are powered by large language models – software trained on vast quantities of text from books, legal documents, websites, and academic sources. Through that training, the model develops a sense of how language works: grammar, logic, argument structure, tone, and genre conventions. But this collection of data lacks the specific refinements and structure that clarity requires. 

To show AI’s capability as an editor, we ran multiple experiments on the three platforms – Gemini, ChatGPT, and Claude. We found the results were not uniform. For example, in response to prompts promoting brevity, some platforms cut too much, deleting essential information, while other platforms retained text that was unnecessary, redundant, or implicit. In one instance a platform added an evaluation not based on the submitted text or prompts and expressly prohibited by a given boundary. In addition, the outputs varied in tone from formal to informal.

Refinement of prompts may be required

You may not draft perfect prompts on the first try. If the first result does not meet your expectations, follow up with a corrective instruction. The model remembers your prior instructions and adjusts based on your corrective instructions.

Effective instructions (prompts)

Lawyers have a choice in how to use AI – either submit to the AI the entire brief and instructions all at once, or submit selected text with fewer specific instructions. Your authors have found that the most precise results are achieved by submitting one paragraph at a time, along with the editing instructions (known as “prompts”).

Effective prompts have three components. 

First, tell the AI who you are, what you are working on, and why.

Second, define precisely the task you want AI to perform. If you have multiple tasks, state them in priority order. If the tasks conflict, the AI will give more weight to the earlier instruction.

Third, provide positive and negative constraints on what AI should and should not do. For example: “Limit each sentence to 12 words.” Or: “Do not alter any citation, defined term, or legal argument.” Another form of constraint is an example of what not to do and what to do: “Not this:…. But this….” Your authors learned the need for constraints when we instructed the AI to shorten an entire brief but, for lack of an appropriate constraint, the AI put all the text in footnotes!

Example of AI applying prompt

     Your authors have tested the three major models – Gemini, ChatGPT, and Claude – to compare the responsiveness of each model to specific instructions. The quality of the AI edits depends largely on the quality of the prompt. Vague prompts produce vague edits; precise prompts produce precise edits.

To test the models’ ability to make text shorter (deleting what is unnecessary, redundant, or implicit), we selected the first paragraph of Justice Cardozo’s opinion in Palsgraf v. Long Island R.R. Co.:

Plaintiff was standing on a platform of defendant’s railroad after buying a ticket to go to Rockaway Beach. A train stopped at the station, bound for another place. Two men ran forward to catch it. One of the men reached the platform of the car without mishap, though the train was already moving. The other man, carrying a package, jumped aboard the car, but seemed unsteady as if about to fall. A guard on the car, who had held the door open, reached forward to help him in, and another guard on the platform pushed him from behind. In this act, the package was dislodged and fell upon the rails. It was a package of small size, about fifteen inches long, and was covered by a newspaper. In fact it contained fireworks, but there was nothing in its appearance to give notice of its contents. The fireworks when they fell exploded. The shock of the explosion threw down some scales at the other end of the platform many feet away. The scales struck the plaintiff, causing injuries for which she sues.” [181 words]

To shorten the paragraph, your authors gave the AI the following prompts: Roles: I am a lawyer for the railroad. I drafted the following document to avoid liability for the railroad. You are a skillful editor capable of condensing wordy text to its essence without any loss of meaning.

• Your task: Delete words that are unnecessary to establish the paragraph’s point, or are redundant, or are implicit. After making deletions, revise as concisely as possible.
• Context: The point of this paragraph is to show that the chain of events leading to Ms. Palsgraf’s injury was so unforeseeable as to bar the railroad’s liability for the guards’ negligence.
• Instructions: Delete every word, phrase, or sentence that describes a fact that was not a cause of Ms. Palsgraf’s injury.
• Boundary: Do not add any interpretation or conclusion.

The three models made the following edits to achieve brevity. We have added our evaluation of the platforms’ success (or failure) at achieving brevity and their conformity to (or violation of) other standards that promote clarity. 

Edits by Anthropic Claude (Sonnet 4.6):

“A man carrying a package jumped aboard a moving train car, and guards helped him board, in the process dislodging his package onto the rails. The package was small, covered in newspaper, and gave no outward indication of its contents, which were in fact fireworks. The fireworks exploded when they fell, the shock of which threw down scales at the other end of the platform, striking the plaintiff.” (68 words)

Evaluation: Each sentence is too long – 25 words; 20 words; 23 words. A prompt was needed to limit sentences to 12 words and revise longer sentences into two sentences. Also, the phrase “fireworks exploded when they fell” failed to put cause before effect (as in the original input): “fireworks when they fell exploded.” 

Edits by ChatGPT 5.2:

“Plaintiff stood on defendant’s railroad platform. As a moving train left the station, two men ran to catch it. One boarded safely. The other, carrying a small package wrapped in newspaper, jumped on and appeared unsteady. A guard on the train reached to help him while a guard on the platform pushed him. The package was dislodged and fell onto the rails. Nothing about the package’s appearance indicated it contained fireworks. When it fell, it exploded, and the blast knocked down scales far down the platform, which struck and injured the plaintiff.” ( 92 words)

Evaluation: Failed to delete unnecessary words, such as: the man who boarded safely, explaining in needless detail how guards helped the man with fireworks board the train, and stating “struck and,” which is implicit. Also, interrupted train of thought by inserting between cause (“falling”) and effect (“exploded”) the implicit information that nothing about the paper-wrapped package’s appearance indicated it contained fireworks. 

Edits by Google Gemini 3: 

“While guards assisted a passenger boarding a moving train, a small package wrapped in newspaper fell from the passenger’s arms. Nothing in the package’s appearance suggested it contained fireworks. However, the package exploded upon hitting the rails, causing a shockwave that dislodged scales many feet away. These scales fell and injured the plaintiff.” (53 words)

Evaluation: The second sentence – that nothing suggested the package contained fireworks – is unnecessary as implicit for a package wrapped in newspaper. “However” at the start of the third sentence lacks any logical relation to prior text. In the third sentence, “the package exploded when hitting the rails” reverses the sequence of time, cause and effect, and climax. In the final sentence, “scales fell” is redundant of “dislodged scales.” Shorter to write: “dislodged scales many feet away, injuring plaintiff.”

The flaws in these outputs show the need for more precise prompts on brevity, as well as on other aspects of clarity.

Here are other examples of how AI can help you edit your briefs.

Task 1: Turn a dense prose paragraph into a list.

Input: “Because of her father’s home auto repairs, Ms. Thomas was exposed to asbestos in the following ways: she assisted with repairs and cleaned up the dust; she played in the driveway next to the garage; she passed through the auto repair zone when entering or leaving the house; she shook and laundered her father’s dusty work clothes; when her father was in the house she was close to his dusty work clothes.”

Prompt: Revise this paragraph into an introductory sentence followed by a numbered list of the ways Ms. Thomas was exposed to asbestos.

Output: “Because of her father’s home auto repairs, Ms. Thomas was exposed to asbestos in these ways:

1. She helped with repairs and cleaned up asbestos-containing dust.

2. She played in the driveway next to the garage where repairs occurred.

3. She walked through the repair area when entering or leaving the house.

4. She shook out and laundered her father’s dusty work clothes.

5. She was near her father’s dusty work clothes while he was in the house.”

(Generated by ChatGPT5)

Task 2: Create a paragraph’s topic sentence.

Input: “After the Civil War, Reconstruction to give Blacks meaningful participation in society was scuttled. In the late 1800s the Supreme Court ruled that racial discrimination in public accommodations was constitutional. The armed forces remained segregated until the 1940s. Racial discrimination in public schools was not declared unconstitutional until 1954. And statutes banning racial discrimination in public accommodations and political districts were not enacted until the 1960s.”

Prompt: Create a suitable topic sentence for this paragraph.

Output: “For nearly a century after the Civil War, law and government repeatedly delayed full civil rights for Black Americans.” (Generated by ChatGPT5)

Task 3: Create parallelism.

Input 1: “Either the court must strike the evidence or grant a mistrial.”

Input 2: “Defendant objects that the statute is inequitable and on the ground of its unconstitutionality.”

Prompt: Express parallel thoughts in parallel form.

Example: Not this: “It was a long closing argument and emotional.” But this: “The closing argument was long and emotional.”

Revise the above sentences to express parallel thoughts in parallel form.

Output 1: “The court must either strike the evidence or grant a mistrial.”

Output 2: “Defendant objects that the statute is inequitable and unconstitutional.” (Generated by Claude and Google Gemini)

Task 4: Shorten a sentence.

Input: “The Hawaii Land Reform Act is classical socioeconomic legislation, designed to serve purposes recognized as vitally public from the very founding of our Republic.”

Prompt: Condense the following sentence to its central point.

Output: “The Hawaii Land Reform Act serves a vital public purpose.” (Generated by Claude & Google Gemini)

Task 5: Simplify a sentence.

Input: “The decree under review protects the petitioner with sedulous forethought against an oppressive Inquisition.”

Prompt: Simplify the foregoing sentence.

Output: “The decree protects the petitioner by carefully guarding against an oppressive investigation.” (Generated by ChatGPT5)

Task 6: Prompt: In the following sentence, replace the nominalization with the embedded verb. Example: Not this: “placed reliance on,” But this: “relied on.”

Input: “The ceiling height is in violation of the city code.”

Output: “The ceiling height violates the city code.” (Generated by Claude & ChatGPT5 & Gemini)

Limitations of AI

AI has well-known limitations, only some of which apply to editing your brief. The ability of AI to hallucinate – make up citations that appear real – will not be a risk in brief editing. Also, though the AI may purport to engage in legal analysis and criticize or recommend legal arguments, AI cannot be relied upon to exercise the legal judgment required of a lawyer. Finally, submitting your brief to an AI platform may raise privacy and confidentiality concerns, which are beyond the scope of this article.

Future possibilities

Editing one paragraph at a time creates precise outputs but is tedious. AI may soon be capable of editing an entire document with no loss of precision. Keep testing your ability to create prompts that are both precise and applicable to the entire document to learn whether the outputs meet your requirements for clarity.

Daniel U Smith is a Certified Appellate Specialist and author of “Persuasive Legal Writing.” www.persuasivewriting.org. 

Roger Ruttimann is a senior software engineer in the San Francisco Bay Area with expertise in building distributed, high-performance, secure software systems.