Guide
How to write a performance review draft with ChatGPT
To draft a performance review with ChatGPT, give it the expectations in your company's own words and dated examples from the whole review period: what happened, what the person did, and the impact. Ask it to rate each expectation separately, to write one to three growth areas as behaviours with a concrete next step, to tie the rating to your scale in a few sentences, and to write the summary last. Then ask for a bias check: examples bunched at the end of the period, one trait colouring every section, vague praise, and personality words. The AI drafts; the manager decides the rating, and only on an account approved for employee data.
Short answer
- Give it: expectations, and dated examples from the whole period.
- Ask for: each expectation rated on its own; growth areas with a next step.
- Bias check: recency, halo, vague praise, personality words.
- The manager decides; approved account only.
The brief to copy
Performance review brief
"Draft a performance review for [role] for [period] from the notes below.
One section per expectation, in our words: [list]. Rate each one separately, from its own evidence, before any overall view.
Write each example as situation, behaviour and impact. Leave out examples without an outcome.
One to three growth areas, each a behaviour with one example and one concrete next step.
A rationale of two to four sentences tying the evidence to our scale: [scale]. Summary last, in three sentences.
Then flag: more than half the examples from the last two months; one trait colouring every section; vague praise; personality words like 'abrasive' or 'bossy'.
Notes: [paste, with no health or family details]"
A worked example
Illustration: a made-up business. Sam manages Priya, an account manager, and pastes his notes for the year. The bias check comes back first: seven of his nine examples are from the last two months, and one note calls Priya "abrasive with the warehouse". Sam adds six examples from earlier in the year, so seven of fifteen are recent, under half, and replaces "abrasive" with what happened. Part of the draft:
From the draft
Client retention (meets the bar): In March, when the largest client's delivery slipped two weeks, Priya set up a weekly call with their buyer and the warehouse lead; the client renewed in June. In October she flagged a pricing error before the invoice went out, which saved a credit note.
Growth area: When a delivery is late, Priya escalates to the warehouse by email with the client copied, which has led to two tense threads. Next step: raise it first with the warehouse lead directly, then update the client with a date.
Rationale: The June renewal and the caught error match "meets" on our scale: she keeps accounts healthy without prompting. The escalation pattern keeps it from "exceeds".
Sam reads every sentence against his notes, sets the rating himself, and keeps the draft in the company's Claude Team account, not a personal one.
Check before you share it
- Every claim has a dated example from your notes.
- Examples cover the whole period, not just the last two months.
- Each growth area is a behaviour with a next step, not a trait.
- No personality words, no vague praise.
- No health, family or other personal details went into the AI; see what not to paste.
Next step: keep the brief
Review season means the same brief for every report. HR prompts, saved shows where to keep it with your company's expectations. For what AI may and may not decide about people, see the HR guide to an employee AI policy.
Skills for this
- ManagementPerformance review draftDrafts a performance review from the manager's evidence, per expectation with examples and impact, then checks it for rating bias and vague words.
- ManagementFeedback with Situation-Behavior-ImpactWrites feedback with Situation-Behavior-Impact plus a request, in positive and corrective versions, sized for Slack or in person.
- Management1:1 prep for managersPrepares a manager's 1:1: the report's agenda first, last time's commitments, one win, one SBI feedback, a growth question, and blockers.
- ManagementCare and challenge feedback checkPlaces planned feedback on Kim Scott's care personally and challenge directly axes, rewrites it to do both, and plans how the manager asks for criticism.
Questions
Can I use ChatGPT to write a performance review?
As a drafting aid, yes, on an account your company has approved for employee data. Give it the expectations and dated examples, ask for each expectation to be rated separately, and check the draft for bias. The manager decides the rating and owns every word.
What should a performance review include?
For each expectation: the evidence, as situation, behaviour and impact; one to three growth areas, each a behaviour with a concrete next step; a short rationale tying the evidence to your rating scale; and a three-sentence summary written last.
How do I avoid bias in a performance review?
Check for the common rater errors: examples bunched at the end of the period (recency), one trait colouring every section (halo or horn), every rating at one end or all in the middle, vague praise, and personality words such as abrasive or bossy. Replace each with a dated example and its impact.
AI skill
Have your AI draft a fair performance review
Put this skill in AI Skill Pod and everyone on your team gets it in their Claude, Cursor or ChatGPT (Pro, Business, Enterprise): one invite link, free to start. Performance review draft has your AI draft each review from the manager's notes: each expectation rated on its own evidence, growth areas as behaviours with a next step, a rating rationale tied to your scale, and a bias check.
14 days free · No card needed
Sources
- Center for Creative Leadership: Situation-Behavior-Impact feedback — clarify the situation, describe the behavior, then share the impact (read 2026-11-05)
- University of Tennessee Health Science Center HR: Beware of rater errors — recency, halo and horn, leniency and severity, central tendency (read 2026-11-04)
- OpenAI Help: Data controls — personal plans train on chats by default unless turned off; read through search-result text (read 2026-11-05)
- Anthropic Privacy Center: commercial products — Claude for Work not used for training by default (read 2026-11-05)