Performance review evidence log

Open paper evidence ledger with reference cards clipped beside entries and an amber tab marking a missing record. — AI-generated editorial illustration.

Keep one row per claim you may need at your next mid-year or year-end review: date, project or area, what happened, outcome or metric, a link or artifact, and who can verify it. At review time, paste the rows into ChatGPT, Claude, Gemini or Copilot with a prompt that drafts the write-up only from those rows and refuses to invent evidence.

The Calderbrook example is made up.

For writing prompts that stay useful at work, see How to write AI prompts for real work.

What columns does the log need?

Use a spreadsheet, Notion table or plain text file. One row is one claim you might put in a review form. Keep these six columns in this order so the prompt can read them cleanly.

Date: when it happened, in day month year.
Project or area: the workstream a manager would recognise.
What happened: one concrete action or event, not a slogan.
Outcome or metric: the measured result, or leave blank if you do not have one yet.
Link or artifact: a ticket, doc, export or folder that still exists.
Who can verify: a named person who saw it, not a team label.

How do you fill a row without inventing evidence?

Add the row on the day the work lands, or as soon as you have the artifact. Copy the date from the ticket or calendar, not from memory a month later. Write what happened in plain verbs. Put a number in outcome or metric only when you can point at the source in the link column. If the result is still open, leave outcome blank and keep the row; blank is honest.

Name the verifier while they still remember. A shared mailbox or "the team" will not help you when the form asks for evidence. Do not pad rows to look busy. Four solid rows beat twelve that rest on vibes.

Example: six rows into a review draft

Here is a short Calderbrook log a delivery lead might keep from January to July 2026, ahead of a mid-year review.

A draft that stays faithful to those rows might read like this.

On Calderbrook migration I cut the finance batch over to the new queue on 14 January 2026; batch error rate fell from 4.2% to 0.8% over four weeks [14 Jan 2026 | Calderbrook migration]. The weekend rollback drill on 3 March restored service in 47 minutes against a 90-minute target [3 Mar 2026 | Calderbrook migration]. In July I replaced the hand-written cutover checklist with a checked runbook; two cutovers used it with no missed steps [18 Jul 2026 | Calderbrook migration].

For client reporting I replaced the monthly slide pack with a live workbook for two sponsors on 11 April 2026; those sponsors stopped asking for a separate pack in May and June [11 Apr 2026 | Client reporting]. On mentoring I coached two juniors through their first production release on 22 May; both released without a severity-1 incident in their first month [22 May 2026 | Mentoring]. On cost control I closed unused staging environments after the Q1 review; £1,140/month hosting came off the June invoice [9 Jun 2026 | Cost control].

That draft invents nothing outside the six rows. If you asked the model for "stakeholder feedback scores" and no row covered them, a faithful answer would say there is no logged evidence for that theme.

What do you check before you submit?

Read the draft against the log line by line. Every claim should carry a date and project or area that exists in a row. Open each link or artifact once. If the model added a metric, stakeholder or project you did not log, delete that sentence. If a blank outcome got a made-up number, put the blank back or fetch the real figure before you submit.

When a claim fails the check, fix the log first if the work was real, then regenerate. Do not polish invented detail into the form.

Related guide

Before you write the prompt for your own form: How to write AI prompts for real work.