The hard part of a self-evaluation is not the writing. It’s remembering what you did in February. Start by dumping every scrap you can find into a list, then use AI for the conversion step: turning “fixed the onboarding flow” into a line that names what changed, what it produced and how anyone could check.
That order is what makes the result usable. Ask a model to write your review from a job title and you get generic praise that any of your colleagues could have submitted. Give it fifteen specific things you did and it will produce fifteen specific claims, in the format your form wants.
Spend twenty minutes gathering evidence first #
Before any prompt, open these and write down anything that was yours:
- Your calendar, month by month. Recurring meetings tell you what you actually owned.
- Sent mail and chat, searched for the months you’ve forgotten.
- Your ticket tracker, pull requests, deal records or case notes, filtered to you.
- Documents you created or substantially rewrote.
- Anything that broke and got fixed, and anything you were pulled into unexpectedly.
Write raw and unpolished. “Covered for R for six weeks”, “cut the report from two days to four hours”, “trained the two new starters”. Fifteen to twenty-five lines is plenty.
If that process hurt this year, the fix for next year is a running note: one document or pinned chat, one line a week, no formatting. Ten minutes a year of effort, and next review you start from a real list instead of an empty box.
The conversion prompt #
Below is a raw list of things I did this review period. For each line:
- Rewrite it as one sentence: what I did, what changed as a result, and
what evidence exists.
- If the line has no measurable result, do not invent one. Mark it
NEEDS NUMBER and ask me the one question that would get it.
- Keep my plain wording. No adjectives like "spearheaded" or "pivotal".
Then group the rewritten lines under the headings my form uses:
[paste your form's headings]
Raw list:
[paste]The “do not invent” rule is the one that matters. Models fill gaps confidently, and a self-evaluation containing a fabricated percentage is a serious problem if anyone checks. Ask for the question instead and answer it yourself, even if the answer is “roughly a third of the team”.
Give each claim a result #
Most weak self-evaluations are a list of activities. Most strong ones are the same list with a consequence attached. The model is good at prompting you for the missing half if you ask it to interrogate each line:
- Activity: “Ran the weekly reporting.”
- With a result: “Ran the weekly reporting and moved it from a manual spreadsheet to a scheduled export, cutting it from about six hours a month to under one.”
Numbers aren’t always available, and inventing them is worse than not having them. Time saved, volume handled, errors avoided, people trained, deadlines met under specific constraints: any of those work, and so does a named person who’d vouch for it.
Writing the development section without sinking yourself #
Nearly every form asks where you need to improve, and both failure modes are common. Write nothing and you look unreflective. Write a confession and it gets quoted back to you.
The shape that works is a real limitation, paired with what you’ve already done about it and what you need. Ask the model for three versions of a development point at different levels of candour, then choose. Something like: “I took on too much direct work in Q2 and delegation suffered. I’ve moved two recurring tasks to the team since and want a clearer escalation rule for the next busy period.”
Read it out loud. If it would embarrass you in a room with your manager’s manager, soften the framing without dropping the substance.
What not to paste into a cloud chatbot #
A self-evaluation draft is one of the more sensitive documents a person handles at work. It typically contains your manager’s name, colleagues’ names and your candid opinion of them, unannounced projects, client names, and sometimes compensation context. Peer feedback about other people is theirs, not yours to upload.
Two rules cover most of it. Don’t paste anything covered by a confidentiality agreement, which is dealt with in using AI with confidential client data. And check whether your employer allows the tool at all, because many organisations permit an approved enterprise assistant and prohibit consumer ones. Using AI at work without leaking confidential data covers how those policies usually read.
Where a document is sensitive enough that you’d rather it not exist anywhere else, Personal LLM runs open models such as Qwen 3.5 and Gemma 4 on the phone itself. There’s no account and no server, so a review draft with your manager’s name in it stays a chat on your device, and you can pin that chat and add to it through the year. It’s free on iOS and Android.
Make it sound like you #
The person reading your review has read your writing for a year. A sudden shift into smooth corporate prose is noticeable, and it reads as distance rather than polish.
Two habits fix it. Paste two paragraphs you wrote earlier in the year and ask the model to match that register. And strike the vocabulary that gives it away: spearheaded, leveraged, robust, seamless, championed, drove alignment. Checking the tone of a message with AI is the same technique applied to anything you send.
Then edit by hand. Cut a third. Self-evaluations are read quickly, usually by someone with twelve of them to get through, and the shorter one with concrete results wins.
Frequently asked questions #
Will my manager know I used AI to write my self-evaluation? #
Not if the content is yours and the voice matches your normal writing. What gets noticed is generic phrasing, a jump in polish, and claims with no specifics behind them. Use the model to structure and tighten your material, not to generate it.
Is it safe to paste my performance review into ChatGPT? #
It depends on your employer’s policy and what’s in the document. Consumer chatbots store conversations against your account, and review material often includes colleagues’ names and unpublished business information. Check your organisation’s rules first, use the approved tool if there is one, and keep the genuinely sensitive drafting on a device you control.
What if I don’t have measurable results? #
Plenty of good work isn’t measurable. Use scope, consistency and difficulty instead: what you owned, what you kept running, what went wrong that you absorbed, who you unblocked. A named colleague who can confirm it is worth more than a made-up percentage.
How long should a self-evaluation be? #
Follow the form’s limit. Where there isn’t one, a page is usually right: five to eight accomplishment lines with results, a short development section, and two or three goals. Longer submissions get skimmed, and skimming favours whatever is at the top.
Can AI help me prepare for the review conversation itself? #
Yes, and it’s a better use than drafting. Ask it to play your manager and challenge each claim, or to generate the three questions you’d least like to be asked. The same roleplay setup works for practising a job interview.