How to write a character letter with ChatGPT or Claude

Step 1: Collect your notes first

AI assistants write best when you give them raw, true material. Before you open ChatGPT or Claude, write down in your own words: who you are, how you know the person and for how long, one specific thing you saw them do, what they have done since the arrest (only what you actually know), and how you will support them. Five short lines are enough.

Step 2: Use a prompt that forbids invention

Copy this prompt, fill in the notes at the bottom, and paste it into ChatGPT, Claude or Gemini. Use placeholders for names while you draft.

Help me draft a one-page character letter for [NAME]'s [sentencing / parole / custody / immigration] hearing. The letter will go to their attorney first.

Rules:
- Use ONLY the facts I give you below. Do not invent stories, feelings, habits, remorse, treatment, accomplishments or promises.
- If something is missing, leave a [placeholder] instead of making it up.
- Write in plain, everyday English that sounds like me — not legal language, no clichés like "pillar of the community".
- Do not claim innocence, argue about the case, blame anyone, minimize what happened, or ask for a specific sentence.
- Do not include addresses, dates of birth, children's names or case numbers.
- Keep it under 400 words, addressed "Dear Judge [LAST NAME]," (or "Dear Your Honor," if I don't know the name).
- After the letter, list every sentence that is not directly supported by my notes.

My notes:
- Who I am (name, city, job):
- How I know them and for how long:
- What they're in court for (one line):
- One true story I saw myself:
- What they've done since (only what I know):
- How I will support them going forward: 

The last rule matters most: asking the assistant to list any sentence it can’t support from your notes makes invented details much easier to catch.

Step 3: The 8 checks before you sign

1. Every sentence is true and first-hand. Delete anything you didn’t personally see or can’t stand behind — especially vivid details the AI added (“he showed up early every day”, “without being asked”).
2. No claim of innocence and no arguing the case. Judges discount letters that dispute the conviction.
3. No blame. Nothing about the victim, police, prosecutors or “the other guy”.
4. No request for a specific sentence. “Please give him probation” is the attorney’s job, not yours.
5. No minimizing. Cut “it was just a mistake”, “nobody got hurt”, “only two beers”.
6. No private details. No addresses, dates of birth, Social Security numbers, case numbers or children’s names — letters often become part of the court record.
7. One page, your voice. Under about 400 words. Read it aloud; if it sounds like a template, rewrite it.
8. Correct address. “Dear Judge [Last name],” — or “Dear Your Honor,” if you don’t know the name. See how to address a judge.

Judges can tell when a letter isn’t really yours

In February 2026, a judge in the Christchurch District Court in New Zealand suspected that a defendant’s apology letters had been written by AI. He typed a short request for a remorse letter into two AI tools himself, found the results matched closely, and gave a smaller reduction for remorse than the defense had asked for (ABA Journal). That case was about a defendant’s own apology, not a character letter from family — but the lesson carries over: a letter only helps when it is clearly in the writer’s own voice and full of details only they could know.

Why invented details are the real risk

AI models are trained to produce fluent, convincing text. When your notes are thin, they fill the gaps with plausible stories, feelings and promises. That reads well — but you are the one signing the letter, and a detail that isn’t true can make the judge doubt everything else.

We measured this while building our own tool. In September 2026 we ran 32 sample cases (including 6 deliberately tricky inputs, such as a writer insisting on innocence or pasting a Social Security number) through two versions of our drafting pipeline, three times each, and had a separate AI reviewer compare every letter with what the writer actually provided. Letters from our earlier, straightforward pipeline contained 13 invented details across the three rounds; in the full test of the version now live, which checks each sentence against the writer’s answers, the reviewer found 0. All 6 tricky inputs were kept out of the final letters. The lesson applies to any AI assistant: checking is not optional.

ChatGPT or Claude vs. a dedicated character letter tool

Either way, the rules come from the same place: published guidance from federal defender offices, such as the Northern District of West Virginia, District of Maryland and Eastern District of Michigan federal defenders. Read one of them before you start — they are short and free.