---
name: performance_review_writing
description: Write, revise, or audit performance reviews, self-assessments, and peer feedback that are fair and usable. Use when drafting a review from notes, turning vague feedback into evidence-based language, choosing or defending a rating, preparing for calibration, or checking a draft for bias, trait language, and unsupported claims.
---

# Performance Review Writing

## Purpose

Produce a review a reader can act on and a rating the writer can defend. Every judgment in the finished review traces to dated evidence from the review period, written in behavior language, rated against standards set in advance, and checked for the rater errors that make reviews unfair.

Three jobs this skill covers:

1. Manager review of a direct report.
2. Self-assessment written by the employee.
3. Peer or upward feedback, and revision of an existing draft.

## Workflow

### 1. Gather the record before writing a word

Collect, do not recall from memory alone:

- Goals and standards agreed at the start of the period, plus the rating scale definitions if the organization has them.
- Work products, metrics, deadlines met or missed, and project outcomes.
- Notes from one-on-ones, feedback given during the period, and the employee's self-assessment.
- Examples spread across the whole period, not only the last few weeks.

If the evidence for a section is thin, say so in the draft and narrow the claim. Never fill a gap with a general impression. A short honest review beats a long invented one.

Stop and ask for more material when: there is no record for most of the period, the rating requested does not match the evidence supplied, or the notes describe the person's character instead of their work.

### 2. Build an evidence table first

One row per goal or competency:

| Area | Standard | Evidence (date, what happened, result) | Gap or strength |
|---|---|---|---|
| Example: Release delivery | Ship agreed scope by date | Mar 14: shipped checkout fix 2 days early, defect count 0 in first week | Strength |

Rate each area separately before forming an overall rating. Rating everything at once is how one strong or weak area drags the rest with it.

For accomplishments, use CCAR: Context, Challenge, Action, Result. In a self-assessment, state your own action with "I" even when the work was shared, then name the team's result separately.

For feedback items, use SBI plus two steps:

- **Situation:** the specific meeting, date, or project.
- **Behavior:** what a camera could have recorded. Actions and words, not motives.
- **Impact:** the effect on the work, the team, a customer, or a deadline.
- **Inquiry:** the question you will ask for their account, when delivering live.
- **Next:** one concrete change, support offered, and a date to check it.

Worked example, original and invented for illustration:

Weak: "Needs to communicate better. Sometimes unresponsive."

Rewritten: "In the April vendor review (Situation), you answered two of the vendor's four written questions and left the other two without a reply for nine days (Behavior). Procurement delayed the renewal by a week and asked me to chase it (Impact). Next: reply to vendor questions within two business days, or send a holding reply naming the date you will answer. We will check this at our June one-on-one."

### 3. Draft in this order

1. **Summary:** 3 to 5 sentences. Overall picture, the rating if one is required, and the one thing that matters most next period.
2. **Strengths with proof:** 2 to 4 items, each with an evidence row behind it.
3. **Growth areas with proof:** 1 to 3 items, each in SBI form, each paired with support and a next step. No growth area appears in a review for the first time if it was never raised during the period. If it was never raised, flag that plainly instead of presenting it as settled.
4. **Goals for next period:** specific, measurable where the work allows it, with dates and a note on how success will be observed.
5. **Development:** one skill to build, how, and what the manager will provide (time, assignment, course, introductions).

Keep the proportions honest. A review that is all praise with a low rating, or all criticism with a high rating, confuses the reader and will not survive calibration.

### 4. Run the language pass

Rewrite every sentence that fails one of these tests. Full lists and rewrites live in `references/phrasing-library.md`.

- **Camera test:** could a camera have recorded it? "Interrupted Ana twice" passes. "Was dismissive" fails.
- **Trait to behavior:** swap labels (lazy, brilliant, difficult, star) for the actions that led you there.
- **Absolutes:** "always" and "never" need a count or they get cut.
- **Vague praise:** "great attitude", "team player", "good job" need an example or they get cut. Praise without proof reads as filler and helps no one in calibration.
- **Double standards:** ask whether you would write the same sentence, at the same standard, for another person in the same role. Check especially comments on style, availability, likeability, and personal life.
- **Off-limits material:** remove rumors, speculation about motives or personal life, family, health, beliefs, background, and any complaint you cannot source. Reviews cover work in the period, nothing else.

Run the linter on the draft when one is available:

```
python3 bin/review_lint.py draft.txt
python3 bin/review_lint.py --strict draft.txt
```

The linter flags patterns, it does not judge the review. Fix or consciously keep each flag and be able to say why.

### 5. Defend the rating

Before finalizing, write a three-line defense:

- The standard for this rating level, in the scale's own words or plain paraphrase.
- Three pieces of evidence from the period that meet or miss that standard.
- The strongest counter-argument (best evidence for the level above or below) and why it does not change the call.

Check your own draft against the common rater errors: halo and horns (one area coloring all areas), recency (last weeks outweighing the period), leniency or stringency (your ratings sit consistently above or below peers for similar work), central tendency (everyone clustered in the middle), contrast (rated against the last person you reviewed instead of the standard), and similar-to-me (rating shared traits as skill). Name the one you are most prone to and state what you did about it.

### 6. Calibration check

If the organization calibrates, prepare a one-page pack per person: rating, the three evidence points, one borderline question for the group. In the session:

- Discuss the highest and lowest ratings first, then borderline cases.
- Challenge evidence, not people. "What did they deliver?" is a fair question. "Are they a culture fit?" is not, unless that term was defined in advance and applied to everyone.
- Change a rating only when new evidence or a shared standard reading warrants it, and record the reason.
- Do not force a distribution. Calibration aligns what a rating means. A quota makes the numbers tidy and the reviews false.

Full session checklist: `references/calibration-checklist.md`.

### 7. Self-assessment variant

Same evidence rules, different voice:

- Lead with outcomes, not activity lists. "Handled 40 tickets" is activity. "Closed 40 tickets, median first reply under 2 hours, no reopened cases" is an outcome.
- Be specific and brief. The reader may be reading many of these.
- Own one real miss: what happened, what you changed, what the later result was. A self-assessment with no miss reads as sales copy.
- Ask for something concrete: a kind of assignment, training, feedback cadence, or tool. Vague asks ("more growth") get vague answers.
- Do not assign yourself a rating unless the form requires one. Supply the evidence and let it argue.

## Output Contract

Deliver, in this order:

1. The review text, ready to paste into the organization's form, with sections matching the form if one was supplied.
2. The evidence table the review rests on.
3. The three-line rating defense.
4. A flags list: anything you cut for lack of evidence, anything the writer should confirm before submitting, and any feedback that is appearing for the first time in this review.
5. For calibration: the one-page pack described in step 6.

If material was missing, the flags list says exactly what is missing and which sections it weakens. Never present a draft built on thin evidence as finished.

## Operating Rules

1. Every factual claim needs a source in the material supplied: a date, a work product, a metric, or a named event. Unverifiable claims get cut or marked as the writer's impression, never stated as fact.
2. Never invent numbers, dates, quotes, incidents, or praise from others. If a metric would help and none exists, write the qualitative result and note that no metric was tracked.
3. Write about behavior and results in the review period only. No diagnoses, no motive claims, no comments on health, family, background, or personal life.
4. Keep ratings and narrative consistent with each other and with pay, promotion, or improvement-plan consequences the writer describes. Flag any mismatch instead of smoothing it over.
5. If the writer asks for language to justify a rating the evidence does not support, in either direction, refuse that part, show what the evidence does support, and offer the honest version.
6. Treat review drafts as confidential. Do not reuse one person's evidence, wording, or rating in another person's review, and do not compare named employees inside a review.
7. Plain language throughout. Short sentences, specific nouns, no corporate filler. If a sentence could appear in any review for any person, rewrite it until it could only describe this person in this period.