The Performance Review Pack

What this is for

You have a dozen reviews due, a self-appraisal you keep avoiding, and one direct report whose feedback has read "needs to communicate better" for two cycles straight. This pack turns scattered notes into specific, defensible writing: for reviews you give, the one you write about yourself, and the conversation you're dreading.

Before you start

• Gather your raw material: notes, Slack threads, old 1:1 notes, last cycle's draft.
• Read the anonymization note below and do that pass before you open any AI tool.

What never goes into a public AI tool

• Full names. Use role tags instead: "the senior analyst," "Employee A."
• Comp, raise amounts, bonus targets, level/band. Add these by hand afterward, in the real document.
• Anything that IDs the company, if the case touches legal risk, a PIP, or termination: client names, codenames, org details.
• Medical, family, or accommodation info. Summarize impact ("attendance was inconsistent"), not cause.
• Anything already in a formal HR/Legal process. Use your company's approved system, or write it yourself.

Do the anonymizing in a separate text file first: names become role tags, dates become relative ("in March"), dollar figures get removed. The AI needs the substance, not the identifiers. (This anonymization pass is the habit worth building for everything you paste anywhere; the One-Page AI Policy turns it into a team-wide rule.)

The six prompts

Prompt 1: Writing a review from rough notes

Use either model.

ROLE: Management writing coach turning raw notes into a formal review.
CONTEXT: Notes below cover [TIME PERIOD]. Role: [ROLE, e.g. mid-level
product designer]. Do not invent accomplishments or issues not in my
notes.
NOTES: [PASTE ANONYMIZED NOTES, e.g. Shipped checkout redesign 2 weeks
late, unclear specs from PM. Mentored two juniors well, unprompted
mentions from both. Missed Q2 planning deadline twice. Strong client
feedback on March presentation.]
CONSTRAINTS: Use only what's given. Flag vague notes as questions
instead of guessing. Balance strengths and growth honestly, no padding.
OUTPUT FORMAT: Sections for Key Accomplishments, Areas for Growth, and a
summary paragraph. Flag anything too vague to use as written.

Prompt 2: Writing your own self-appraisal with evidence

Use either model.

ROLE: Helping me draft my self-appraisal for this cycle.
CONTEXT: Role: [ROLE]. Goals this cycle: [LIST 2-4 GOALS]. Evidence
below, in no order.
EVIDENCE: [PASTE, e.g. Led the migration, under budget. Onboarded new
hire faster than usual. Missed June deadline due to vendor delay,
recovered by July. Presented to the board twice.]
CONSTRAINTS: No inflated verbs ("spearheaded," "revolutionized").
Plain language tied to outcomes. Keep accountability on misses, but
state what I controlled and what I didn't. Tie each item to a stated
goal.
OUTPUT FORMAT: One paragraph per goal with evidence, plus a closing
paragraph on next-cycle focus.

Prompt 3: Turning vague feedback into behavioral examples

Use either model.

ROLE: Converting vague feedback into specific, observable examples.
CONTEXT: Vague version: [THE VAGUE FEEDBACK, e.g. "Needs to communicate
better with stakeholders"]. Fragments I remember: [WHAT YOU RECALL,
e.g. Didn't loop in Finance before April's pricing change. Sent a May
status update to the wrong list. Didn't flag a delay on the June client
call until asked.]
CONSTRAINTS: Turn each fragment into a dated, observable behavior, not
a trait. No "always/never" unless the fragments show a real pattern. If
I've given only one example, say so before calling it a pattern.
OUTPUT FORMAT: List in situation / behavior / impact format, ready to
read aloud.

What the transformation looks like. In: "Needs to communicate better with stakeholders." Out:

In April, when the pricing change was decided (situation), Finance was not looped in before it went out (behavior), which meant they first heard about it from a client and had to rebuild the quarter's forecast (impact).

That's the difference between feedback someone can argue with and feedback someone can act on. One is a character judgment; the other is a correctable event with a date on it.

Prompt 4: Prepping for a review you expect to go badly

Use either model.

ROLE: Blunt, experienced HR business partner prepping me for a hard
review conversation.
CONTEXT: Delivering to [ROLE/RELATIONSHIP, e.g. a direct report who's
been defensive before]. Core message: [ONE SENTENCE, e.g. Output has
been inconsistent, two deadlines missed with no early warning].
Expected reaction: [WHAT YOU EXPECT, e.g. will argue deadlines were
unrealistic].
CONSTRAINTS: Give me the opening line, not a full script. Anticipate
the most likely pushback and give one honest response, not a
deflection. Don't suggest I soften the core message.
OUTPUT FORMAT: Opening line, pushback plus my response, three questions
to ask before I leave the room.

Prompt 5: Writing a promotion case

Use either model.

ROLE: Building a promotion case document.
CONTEXT: Proposing [ROLE TAG] for promotion from [CURRENT LEVEL] to
[TARGET LEVEL]. Evidence of next-level work below.
EVIDENCE: [PASTE, e.g. Took over vendor negotiation solo, saved 12% on
contract. Two peers requested to work on their projects specifically.
Mentored a struggling teammate back to good standing in one quarter.]
CONSTRAINTS: Map each item to a competency expected at the target
level; don't just list achievements. Flag any thin evidence so I can
find more before submitting.
OUTPUT FORMAT: One page: Scope of Impact, Evidence of Next-Level Work,
Identified Gaps.

Prompt 6: Documenting a performance problem defensibly

Use either model, and read the legal note below first.

ROLE: Meticulous documentation writer. This needs to hold up if it
becomes a formal process.
CONTEXT: Factual record only. Role: [ROLE TAG]. Timeframe: [DATE
RANGE]. Incidents: [PASTE, e.g. March 3: missed client deliverable,
no advance notice. March 20: verbal warning given, documented in 1:1
notes. April 15: missed a second deadline, same project.]
CONSTRAINTS: State only what happened and when. No motive, no
character judgments. Note where feedback or warnings were already
given. Flag any missing date or witness as a weak point.
OUTPUT FORMAT: Dated incident log, then a short pattern summary,
written so someone else could defend it in a meeting.

The legal note, and it's not boilerplate. Using AI in employment decisions is specifically regulated in a growing list of places, and the list is moving fast enough that you should check your own jurisdiction rather than trust any guide, including this one. As of September 2026: six states plus New York City have laws aimed at AI in employment decisions, four of them already in force. Illinois amended its Human Rights Act effective January 2026 to cover AI in discipline and discharge, with an employee notice requirement, though the agency rules defining what that notice must say were withdrawn and have not been refiled. New York City's Local Law 144 requires an annual bias audit and advance candidate notice for automated employment decision tools. Colorado is the cautionary tale about relying on a summary: its 2024 AI Act was repealed before it ever took effect and replaced by a narrower law that starts January 2027. In the EU, employment uses are classified high risk, but those obligations were deferred to December 2027; the separate transparency duties are already in force. Some employers prohibit AI in employment decisions outright. So: before anything from Prompt 6 feeds a PIP, discipline, or termination, loop in HR, know whether your company has an AI-in-employment policy, and treat the AI's role as formatting your factual record, never deciding or recommending the outcome. The decision, and the accountability for it, stays with you.

Getting it right

• Strip names and comp first. Build the anonymize pass into your workflow. Do it in a separate file before opening the AI tool at all.
• Fill thin notes with facts, not invention. Feed more detail, or answer the "flag as a question" prompts instead of ignoring them.
• Edit before it goes into the official system. These are strong first drafts, not finals. Read each one like you're editing someone else's writing about someone you actually know.
• Make your own review sound like you. If your direct report can tell AI wrote it, so can your boss. The fix is the same as everywhere else: the AI formats your specifics; it doesn't supply them. A review with real dates and real incidents can't sound generic.

The 2-minute version

Anonymize first: names to role tags, dollars out, dates relative. Pick the matching prompt. Paste real notes into the brackets. For anything that could become a formal process, loop in HR before AI touches it. Edit the output before it goes anywhere official.

---

More where this came from. The resource library has 24 free guides like this one, written for people with real careers rather than developers.

Weekly, on Tuesdays. The GenXcelerate newsletter: practical AI intelligence for people who already know how to do the job. Subscribe free.

Experience Is the API™. GenXcelerate