AI PR descriptions that sound like you

I started Thirdraft because of code review.

We had a contractor who answered review comments by pasting them into Cursor and posting whatever came back. The replies were polite, complete, and obviously AI. After a while the reviewers stopped trusting anything that looked AI-written. Not just his replies. Anyone's.

Then I caught myself doing the same thing. I'd debug something carefully, write up what I found, and paste it into ChatGPT to clean it up. What came back looked vibe-coded. The work behind it was real. The write-up made it look like it wasn't.

Why it matters in code review

A PR description is a claim: here's what changed, here's why, and I checked it. Reviewers read the tone as a sign of how much care went in. When it reads like a model wrote it, they assume a model decided what goes in it too. Then they either review twice as hard or skim.

Your reviewers are probably good at spotting it. An ACL 2025 study found that people who use LLMs often are better at spotting AI text than most detectors. That's most engineering teams now.

What gives it away

• Headers and bullet lists on a three-line change.
• Openers like "This PR introduces" or "This change enhances".
• A summary that repeats the title.
• Words like "robust", "comprehensive" and "streamline".
• A testing section listing checks nobody ran.

Set it up in Cursor

1. Click Add to Cursor on the MCP page, or add Thirdraft to ~/.cursor/mcp.json yourself.
2. Sign in to Thirdraft in the browser window that opens, and click Allow.
3. Give it your voice. Paste in 3 to 5 PR descriptions, review replies or bug write-ups you wrote yourself, and ask: "Add these to Thirdraft as writing samples." You can also add them at thirdraft.com/app. Real PRs make the best samples, because Thirdraft picks the samples closest to what you're writing.
4. Add a rule so Cursor uses it without being asked (Cursor's rules settings, or a file in .cursor/rules/):

When you write a PR description, a reply to a review comment or a bug write-up for me, run it through Thirdraft's rewrite_in_my_voice tool with context like "PR description" and use the result.

Set it up in Claude Code

1. Run this once:

claude mcp add --transport http thirdraft https://api.thirdraft.com/mcp

2. Type /mcp in Claude Code and sign in to Thirdraft.
3. Add samples the same way. Paste a few PRs you wrote and ask Claude Code to add them to Thirdraft.
4. Add the rule to your CLAUDE.md:

When you write a PR description, a reply to a review comment or a bug write-up for me, run it through Thirdraft's rewrite_in_my_voice tool with context like "PR description" and use the result.

Then ask for a PR the way you normally would:

Write the PR description for this branch from the diff and the commits. Keep it short, only mention tests I actually ran, then run it through Thirdraft and open the PR.

What it fixes and what it doesn't

Thirdraft fixes how the description sounds. It doesn't make it true. If your agent misread the diff, you get the same misreading in your voice. So read your own diff. Check that the description says what the code does, and delete anything you didn't do, like tests you didn't run. Then let it sound like you.

It's told to keep facts, names and numbers as written. Still, glance at file and function names in the result before you post it.

Sources

• ACL 2025: People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text

Questions