Self-review drafting: interview-first prompting, no fabricated metrics, voice matching

Step 1: Dump your notes, messy is fine

Before you open the chat, spend ten minutes collecting what you actually did. Look at:

• Your calendar for the year.
• Emails you sent that you were proud of, or that thanked you.
• Projects, problems you fixed, people you helped.
• What was hard, and what you'd like to learn next.

Write it as a list. Half sentences are fine. If your company has rules about AI tools, check them, and leave out anything confidential.

Step 2: Let it interview you first

Paste the notes and send:

Here are my rough notes for my self-review. Before you
write anything, ask me three questions that would help you
understand what I did and why it mattered. Wait for my
answers.

The questions are the useful part. They pull out the details a review needs, like what changed because of your work, and you answer them in your own words.

Step 3: Set the rules, then let it write

After you answer, send:

Now write my self-review in first person, in plain
language. Use only what I told you. Never invent numbers:
write [ADD NUMBER] where a number belongs. Don't use these
words: spearheaded, leveraged, synergy, drove impact,
passionate. Keep each section short.

Step 4: Make it true and make it yours

• Replace every [ADD NUMBER] with a real number, or delete the sentence.
• Read it out loud. Cut anything you'd be embarrassed to say to your manager's face.
• Keep one honest line about what didn't go well and what you're doing about it. That reads as confident, not weak.

If the draft flatters you too much, ask it to point out where you're overstating. Here's how to get ChatGPT to stop agreeing with you.

What not to hand over

Your review is about you, but your notes may mention coworkers, customers, or company numbers. Leave out names and anything you wouldn't share outside work. You can add specifics back in your own document.

Under the hood

Generic "corporate" phrasing is the model's high-probability default for the genre: performance reviews in its training data are full of it. Three controls shift the distribution:

• Interview-first adds specific, low-frequency detail from you, which crowds out filler.
• Placeholder metrics block a known failure: models produce plausible percentages to satisfy "quantify your impact" patterns.
• Negative style constraints (a banned-phrase list) are blunt but effective. A short sample of your own writing works even better as a positive constraint.

None of this checks truth. The last pass, replacing or deleting every placeholder and reading aloud, is the control that matters.

Short close

Paste rough notes, let it ask questions first, ban made-up numbers, and cut the buzzwords. Fill in the real numbers yourself. Every line should be something you can stand behind.