Draft factual performance-review notes without adding judgements

Example result.

Rowan wrote and shared a supplier-onboarding checklist in February. Two colleagues reported using it. [14 Feb, checklist]
The first draft of the quarterly capacity report was three days late in June. We discussed it and agreed Rowan would start the next one a week earlier and flag slippage before the deadline. [21 Jun, capacity report]
Not established by these notes: agreed development objectives.

The assistant does the writing here. It doesn’t do the assessing. This workflow organises and words observations; it must not produce ratings, judgements about suitability, pay or promotion recommendations, or decisions about capability, discipline or redundancy. Those need the authorised human process.

1. Select the evidence before opening the chat

Gather dated observations across the whole review period and the objectives they relate to. A dated note is still someone’s account, not a complete or unbiased record of the year.

Remove health, pay, complaints, disciplinary material and other sensitive personal information before you paste anything. Leave out confidential feedback from colleagues unless you have permission to use it. Replacing a name with a letter doesn’t make a distinctive situation anonymous.

Give each note a short reference, such as its date and heading. Keep observations, your interpretations and agreed next steps separate. “Report arrived three days after the agreed deadline” describes an event. “Unreliable” is a judgement with much wider implications. Don’t ask the assistant to turn the first into the second.

Why this works: the draft can only be as factual as the notes. Separating events from interpretations now stops the model blending them later.

What you should see: a short, referenced list of job-related notes with nothing sensitive in it.

2. Set the writing task and its limits

Open a fresh chat in your work account. Paste only the selected notes; this doesn’t need access to your HR system.

Draft factual narrative under these review headings using only my notes below. This is a wording task for human review, not an assessment.

Preserve dates, scope, uncertainty and who did the work. After each factual sentence, give its note reference. Do not invent achievements, causes, intentions, character traits, future objectives or agreement.

Do not turn one or several examples into a claim of a consistent pattern. Describe the specific examples instead. Keep an observation separate from an interpretation and from an agreed next step. For anything that went wrong, say what happened, what was discussed and what was agreed, and nothing more.

Do not score, rate, rank, compare people, recommend an employment outcome, or infer suitability, personality, pay or promotion. Leave any assessment field blank.

End with "Not established by these notes": headings or points the notes give no evidence for.

Review period and headings: [period and headings]
Notes with references: [notes]

[period and headings] is the review period and the form’s headings; [notes] is your referenced list from step 1.

Why this works: sentence-level references let you spot a paragraph that starts faithfully and then adds a conclusion. The “Not established” list stops an empty heading being filled with appraisal language because the form looks unfinished. It’s a coverage list, not a statement that the person lacks an ability.

What you should see: short paragraphs under each heading, a reference after every factual sentence and a “Not established” list at the end.

3. Check each statement against what was recorded

Read the draft beside the notes, one sentence at a time:

• Attribution: did this person do the work, contribute to it or report someone else’s?
• Scope: did one report become every report, or one meeting a sustained pattern? Two examples establish two examples, not consistency.
• Hedges that have hardened: your note said “seemed more confident”; the draft says “chairs meetings confidently”. That’s a different claim, and the most common way these drafts drift.
• Outcome and agreement: was a suggestion promoted into an improvement, or a proposed step into something agreed?
• Praise you didn’t write: models are fluent in appraisal language and will produce “a valued member of the team” unprompted. Unearned praise isn’t harmless: it makes the critical paragraph read as the exception, and next year’s honest review look like a sudden decline.

Keep context the notes contain: if a task was late because an input arrived late, dropping that changes the account. Expect to rewrite about a third of it. The draft is a scaffold for your judgement, not a document to approve.

Why this works: each check targets a specific way factual wording turns into assessment.

What you should see: a marked-up draft with overreaching sentences flagged and their notes beside them.

4. Bring the draft back into the human review

Use the draft to prepare for the review conversation, alongside the full authorised evidence. Resolve gaps with the right people, and if you fill one from memory, say it’s a later recollection rather than a note from the time.

Acas describes reviews as discussions about work, objectives, support and development, and advises keeping a written record of the discussion and sharing it with the employee. The AI draft isn’t that record: the conversation may correct it or add agreed next steps.

Follow your organisation’s retention rules for notes, drafts and the final record: don’t delete working material under a legal hold or dispute, and don’t leave copies in a shared project.

Why this works: the draft feeds a human process; it doesn’t replace it.

What you should see: a checked factual draft ready for the reviewer, with open points listed.

Check the result

• Every reference in the draft exists in your notes and says what the sentence says.
• No sentence states a pattern, trait or rating, and assessment fields are blank.
• Anything that went wrong reads as what happened, what was discussed and what was agreed.
• The “Not established” list is there and matches the real gaps.

The usual failure: a specific example becomes a general quality. “Covered three stand-ups” becomes “runs concise meetings”; a late report becomes “needs to improve time management”. Spot it by looking for words such as consistently, always, struggles with or reliable, then checking how many notes sit behind them.

This sentence goes beyond the notes: [sentence]. The note says: [note text and reference]. Rewrite it to keep that scope and uncertainty. Do not replace it with another assessment. Check nearby sentences for the same overstatement and list any you find.

[sentence] is the overreaching line; [note text and reference] is the exact note it should rest on. Check the revision yourself. Asking an assistant to be fair, repeatedly, doesn’t establish fairness.

Fictional example.

14 February 2026, supplier checklist:
Rowan wrote the supplier onboarding steps as a one-page checklist after a repeated sequencing error. Shared it in the team channel. Two colleagues reported using it. The note does not say who asked Rowan to prepare it.

9 May 2026, meeting cover:
Rowan covered three Thursday stand-ups. Seemed more confident by the third. The note records a fifteen-minute duration for the final session only.

21 June 2026, capacity report:
First draft of the quarterly capacity report was three days late. Discussed. Agreed to start it a week earlier next quarter and to flag slippage before the deadline rather than on it.

2 September 2026, handover:
Rowan wrote a handover note for a new coordinator without being asked. The new starter referred to it during their first week.

Headings: Documented contributions; Meeting cover; Deadlines; Agreed future steps.

The first draft opened: “Rowan consistently takes initiative and runs concise meetings, though they need to improve their time management.”

• The failure. All three claims overstate the notes. “Initiative” rests on September only; the February note doesn’t say whether the checklist was requested. The fifteen minutes applies to one stand-up. One late report became a trait.
• The repair. The repair prompt, run on each claim, gave: “Rowan covered three Thursday stand-ups in May. The manager’s note says Rowan seemed more confident by the third; the final session lasted fifteen minutes. [9 May]” and, for June, the paragraph in the example result.
• Why the June paragraph works. It doesn’t characterise and doesn’t say “needs to improve time management”. It records what happened, what was discussed and what was agreed — the shape of a defensible note about a missed deadline. “Agreed future steps” stays in the “Not established” list apart from the June agreement, because nothing else was recorded.

Task kit for this job: Download (ZIP) · Reusable instructions

Go deeper

Map these notes by month and review heading. List note references only, then identify gaps in coverage. Do not rate performance, compare people or infer what happened during a gap.

Sources

• Acas: Performance management
• Anthropic Privacy Center: Is my data used for model training?
• OpenAI: Security and privacy