用AI寫績效評語：6個提示詞與一份範例評語

提示詞帶入目標、你自己的證據和評分等級，評分由你決定

當一個提示詞帶有三樣東西時，就能寫出可用的績效評語初稿：週期開始時訂下的目標、你自己從工單系統、行事曆和數字裡整理出來的證據，以及附有各級定義的評分等級。摩根大通正是允許主管這樣做，並訂下準則：這個工具「不能取代人的判斷」，同時禁止用AI來給分或決定薪資與升遷（Bloomberg，2025年11月13日）。以下是我們在2026年9月29日送給兩個模型的提示詞；其中的客服工程師和她的數字都是虛構的，重點在於這個格式。

Write the manager's comments for an annual performance review. Use only the evidence below; where you need a fact I have not given, write [NEEDED] instead of inventing it. Structure: one paragraph per goal (outcome first, then the evidence), one paragraph on areas to develop, then a proposed overall rating on our scale with one sentence of justification. Under 250 words. No adjectives about the person's character, only what they did and what it produced.

Role: support engineer, 2 years in role. Rating scale: 1 Below expectations, 2 Partially meets, 3 Meets, 4 Exceeds, 5 Far exceeds. A 3 is the expected rating for a solid year; 4 needs measurable impact beyond the role.

Goal 1: Cut median first-response time on tier 2 tickets from 9 hours to under 4. Evidence: Zendesk report, median first response 3.6 hours for Q2 to Q4; 3,100 tickets handled in the year against a team average of 2,400.

Goal 2: Ship the customer-facing status page. Evidence: shipped 14 March, six weeks after the 1 February target; the delay was a vendor SSO dependency outside her control; 22 incidents posted to it since launch.

Goal 3: Mentor two new hires to unsupervised ticket handling within 8 weeks. Evidence: both new hires handling tickets alone by week 6 (manager's one-on-one notes, April and May).

Development note: two escalations in Q3 where the customer was told a fix date before engineering had confirmed it (Jira SUP-2211, SUP-2287).

這個提示詞裡有兩句指示發揮了大部分作用。「[NEEDED] instead of inventing it」（缺少就寫 [NEEDED]，不要自己編）能防止模型在目標上亂加你從沒量過的數字：在我們的實測中，Claude Sonnet 4.5在我們沒提供兩次升級處理後改善證據的地方寫了一個 [NEEDED]，而Gemini 2.5 Pro一個都沒寫，還加了一句我們從沒提供過的流程說明。「No adjectives about the person's character」（不要用形容人格特質的形容詞）則讓初稿只談實際發生的事，這也是員工可以提出討論、人資可以站得住腳的部分。

證據必須由你提供，因為模型手上沒有別的東西。Lattice的2025年人才策略現況報告（State of People Strategy Report）發現，49%的主管覺得很難回顧一整年的回饋，而這一整年的資料正是提示詞需要的：Zendesk或Jira匯出的檔案、行事曆、提交紀錄、一對一面談筆記、儀表板上的數字。提示詞2會在寫任何東西之前，先把這些匯出資料整理成證據行。

Below are raw exports covering this person's review period: [paste the ticket list, the calendar export, the commit log or the sales report, one after another]. Turn them into evidence lines, one per line, each naming a date or date range, a thing that was done and a number from the export. Group them under these goals: [paste the goals]. Where an item supports no goal, put it under "Other". Where a number is missing, write [NUMBER NEEDED]. Do not summarize or rate anything, and do not add any item that is not in the exports.

你並不算早。Resume Builder在2025年6月調查了1,342位有直屬部屬的美國主管：65%在工作中使用AI工具，其中91%用它來評估績效，而ChatGPT是53%的人的首選。

用AI寫自評：同樣的輸入，加上一個誠實的失誤

用AI寫的自評從同樣三項輸入開始，而且只為一個評分辯護：你的證據所支持的那一個，並用平實的話說出你會做得不一樣的一件事。沒有任何失誤的自評，讀起來像是機器寫的，或是寫的人不夠用心，而且已有26%的員工懷疑自己收過AI寫的評語（ZeroBounce，對1,000位美國專業人士的調查，2025年9月）。花旗推出的順序顯示這件事有多普遍：從2025年6月起，員工就能用公司的AI工具起草自我評估，四個月後主管也獲得同樣的協助。

Write my self assessment for the [period] review. My goals were: [paste]. My evidence, all from my own records: [paste the evidence lines from prompt 2]. Our rating scale: [paste, with the definition of each level]. First person, under 300 words, one paragraph per goal that leads with the result and then the evidence, then one paragraph naming the thing I would do differently: [describe the miss in one line]. Argue for the rating my evidence supports and say which level it is; do not argue for a higher one. Where you need a number I have not given, write [NUMBER NEEDED]. No claims about my attitude, effort or potential, only what I did and what it produced.

自評裡的每個數字，都應該是你的主管能在你取用資料的系統中查證的，因為他們就會去那裡看。撐起一份自評的三項成果，也正是撐起一份加薪請求的三項成果，所以用AI爭取加薪裡的論述提示詞，可以替兩者產出要點，而自評是先把它們記錄下來的更好地方。

最後一關是語氣，而且它在這裡比在主管初稿裡更重要，因為讀起來像模型寫的自評，等於告訴你的主管這不是你自己寫的。提示詞4用你自己的句子當範本，對任何工作電子郵件也是同樣的做法。

Here are three paragraphs I wrote myself, from emails or documents: [paste]. Rewrite the self review below so it matches how I write: my sentence length, my vocabulary, the way I start sentences. Remove every adjective that is not attached to a number. Do not add, remove or reorder any fact, and keep every number and every [NUMBER NEEDED] exactly as it is. [paste the draft]

寫績效評語哪個AI最好？我們用同一個提示詞跑了兩個

Claude Sonnet 4.5寫出較謹慎的評語，Gemini 2.5 Pro寫出較溫暖的，而兩者從相同的證據得出同一個建議評分：4分。我們把提示詞1原封不動地透過同一個API（fal.ai的any-llm端點）送給兩個模型，時間是2026年9月29日，temperature設為0.3，各跑一次，並開啟Gemini的推理功能，因為該端點對這個模型要求如此。整個測試花費不到一美分。

Claude的輸出，除了粗體標記之外未經編輯：

Goal 1: First-Response Time The median first-response time target was met, with Zendesk data showing 3.6 hours across Q2-Q4, below the 4-hour goal. This was achieved while handling 3,100 tickets, 29% above the team average of 2,400.

Goal 2: Status Page The customer-facing status page shipped on 14 March, six weeks later than the 1 February target due to a vendor SSO dependency outside her control. Since launch, 22 incidents have been posted to the page.

Goal 3: New Hire Mentoring Both assigned new hires reached unsupervised ticket handling by week 6, two weeks ahead of the 8-week target, as documented in manager one-on-one notes from April and May.

Development Areas Two escalations occurred in Q3 (SUP-2211, SUP-2287) where fix dates were communicated to customers before engineering confirmation. [NEEDED: evidence of improvement or corrective action taken after these incidents.]

Proposed Rating: 4 (Exceeds) The combination of meeting the response-time goal while handling significantly above-average ticket volume, accelerating new hire readiness, and delivering the status page despite external delays demonstrates measurable impact beyond standard role expectations.

Gemini的輸出，未經編輯：

You exceeded the goal to reduce median first-response time, achieving 3.6 hours against a target of under 4 hours from Q2 to Q4. This was accomplished while handling 3,100 tickets for the year, above the team average of 2,400.

The customer-facing status page was shipped on 14 March. While this was after the 1 February target, the delay was due to a vendor dependency outside your control. The page has been used to post 22 incidents since launch.

You met the goal to mentor two new hires. Both were handling tickets unsupervised by week 6, ahead of the 8-week target, according to one-on-one notes from April and May.

For development, focus on communication during escalations. In Q3, two customers were given a fix date before engineering had confirmed it (SUP-2211, SUP-2287). The process is to wait for an engineering estimate before providing a date to the customer.

Proposed rating: 4 (Exceeds). The significant outperformance on ticket volume and response time created a measurable positive impact on the customer experience beyond the defined expectations for the role.

評分是兩者傾向一致的地方，也是評分必須留在你手上的原因。量表說3分是紮實一年的預期評分，4分則需要超出職務範圍、可衡量的影響。兩個模型都把3,100張工單對上平均2,400張，當成那樣的影響，而且都沒有在評分的那句話裡，把目標2的六週延誤或兩次升級處理拿來衡量。以這份證據來看，4分說得通；3分再加一行談升級處理的內容也說得通，而模型兩次都選了比較寬厚的那個。提示詞5會讓它在你做決定之前先為另一邊辯護。

Here is a draft review and our rating scale: [paste both]. The draft proposes a [rating]. Write the strongest one-paragraph case for the rating one level below it and the strongest case for the rating one level above it, using only the evidence in the draft and the scale's own definitions. Then list every piece of evidence the proposed rating ignores. Do not recommend a rating.

所以，哪個AI最好：對評語來說，模型的影響比不上證據，因為兩個一線模型拿到同樣的證據，得出同樣的評分，差別只在語氣和一句捏造的話。想要它指出缺少的證據就選Claude，想要初稿直接對員工說話就選Gemini。把兩者放在同一處有幫助的情境，是有八位部屬的考核季。在Whizi上，你可以把評分等級和能力定義一次釘選在一個專案裡（截至2026年9月，一個專案最多可釘選10個檔案、每個10 MB，包含.docx和PDF，專案指示最多32,000個字元，會隨每則訊息一起送出），然後把每位部屬的證據分別交給Claude和Gemini，不用再貼一次量表（專案如何運作）。Whizi輸的地方是工作的另一半：它沒有HRIS整合，也沒有團隊帳號，所以不會替你從Workday、Lattice或Zendesk拉資料。花旗的Performance Assist「automatically gathers all of that information directly from our systems」（直接從我們的系統自動收集所有這些資訊），這是花旗薪酬與福利主管Patricia Gould的說法，而Lattice的AI Agent則在Lattice內部彙整過往的回饋。在Whizi上，你要自己匯出再貼上。

你必須告訴人資或員工，這份評語有AI協助撰寫嗎？

在兩家有公開立場的大型雇主，答案並不相同：花旗會告知員工，並讓任何人選擇不要AI起草的評語，到2025年12月，選擇退出的人不到1%；而摩根大通的準則，依Bloomberg的報導，是限制主管能用這個工具做什麼，並不要求向部屬揭露。摩根大通禁止由AI給分；花旗則讓主管對最終評估負責。其他地方的答案，要看貴公司的AI政策，而Lattice和Textio的第一項檢查清單，和我們的一樣：貼上之前，先讀過政策。以下是2025年11月13日至12月2日期間報導的、有具名的美國雇主立場：

調查數字顯示，這個問題在大多數團隊裡已經存在。在Resume Builder 2025年6月的調查中，「More than 20% say they allow AI to make decisions without human input either all the time (5%) or often (16%)」（超過20%的人表示，他們讓AI在沒有人為輸入的情況下做決定，不是一直如此（5%）就是經常如此（16%）），而在用AI管理人員的主管中，只有32%接受過任何正式的AI訓練。ZeroBounce 2025年9月的調查則顯示，41%的主管曾用AI起草或修改評語。Gartner在2025年12月對1,622位受訪者的調查發現，「managers reported an average of four hours saved across different parts of the performance management process when using AI」（主管表示使用AI時，績效管理流程的各個環節平均省下四小時），而它在2026年4月的建議是把AI當作「as an input to managerial judgment rather than a replacement for it」（管理判斷的輸入，而不是取代它）。

關於揭露的唯一一項實地實驗，發現告知別人有代價，也找到了繞過的方法。Tong、Jia、Luo和Fang於2021年在Strategic Management Journal上，衡量了一家真實公司內部由AI產生的回饋，並寫道「employees may develop a negative perception of AI feedback once it is disclosed to them, thus harming their productivity」（員工一旦得知回饋來自AI，可能對它產生負面觀感，進而損害他們的生產力），而這個效應會隨員工年資增加而減弱。他們建議的解法，正是上述每項政策最後落腳的形式：不論是什麼起草的，回饋都由主管用主管自己的語氣傳達。CM Law的合夥人Peter Cassat在2025年12月告訴HR Dive，用AI評估員工「could subject organizations to discrimination claims as well as substantial compliance obligations」（可能使組織面臨歧視訴訟以及重大的合規義務），這就是銀行讓評分留在人手上的法律原因。各州與各城市的規定不同，本頁也不是法律建議；該問人資或法律顧問的問題是：在你所在的地方，起草工具算不算自動化決策工具。人資用AI列出六項美國的AI就業規定，包含涉及薪資與懲處的規定，以及在自己的條文中把起草排除在外的規定。

我可以免費用AI寫評語嗎？

可以。ChatGPT、Claude和Gemini的免費方案都能直接使用提示詞1，而專用的產生器做的是同一份工作，只是用表單取代提示詞。TechnologyAdvice在2026年3月測試了八個免費產生器，保留了四個：Easy-Peasy.AI（每天2份評語）、Writify.AI（無限制，含廣告）、Vondy（註冊前4份）和GravityWrite（每月1,000字）。這些表單會要求填員工姓名；對任何你還沒讀過資料條款的工具，請改填職稱（AI對話中絕對不能貼的內容）。

不論免費或付費，都要做一項表單工具會略過的檢查。Textio在2023年1月對ChatGPT回饋的測試發現，它在10次裡有9次預設護理師是女性，而且為女性寫的回饋約長15%，多出來的長度大多是批評。提示詞6會在送出任何東西之前，對你自己的那一批評語做這項檢查。

Here are [N] draft reviews from my team, with names replaced by letters: [paste]. Compare them and list: any review more than 20% longer or shorter than the median; any adjective about character (reliable, difficult, bubbly, aggressive) and which review it appears in; any pronoun or gendered word; any claim with no number or date behind it. Output a table with one row per finding. Do not rewrite anything.

本頁每項政策都同意的規則，一句話就說完了：證據由你提供，初稿可以來自模型，評分與面談則始終是你的。