Manager tools

Quarterly check-ins and annual reviews, without the blank page.

Keep a light record through the year — goals, a ten-minute check-in each quarter, what you owe them. When a review is due, answer a few guided questions and get a written draft to edit. You set the rating. It never does.

The people you write about never need an account.

How it works

Three parts, in the order they happen.

1 · Through the year

Keep the record

Goals for the year, each with a measure and a target. At the end of every quarter, a short check-in: where each goal landed, what went well, what got in the way, and a band for the quarter. Ten minutes, while you still remember it.

2 · When a review is due

Answer the guided questions

4 questions for a quarter, 7 for the year, most with a tick-list to jog your memory before you type. Notes are enough. In either, you can ask up to ten colleagues three written questions — a couple of people for a quarter, more for a year.

3 · The document

Edit the draft, set the band, export

Your notes, the record and any colleague input become a document a person can read. Every sentence is editable. The band is yours to set with a reason beside it, and the whole thing downloads to hand over.

Worked example

Watch one being written.

Two invented people, one real product. The questions, the tick-lists and the documents are the ones the app uses, rendered by the same code, and both drafts pass the app’s own checks in our test suite. Pick a period and step through what went in and what came out.

Q3 2026 · Ravi, Customer Support Lead

Four questions and two colleagues, both of whom say the same thing the record already says. The short one that gets written four times a year.

Step 1 of 5

What was already recorded about Ravi

Kept through the year, a quarter at a time. Every number the draft uses has to come from here — it may not use a figure that was never recorded.

Ravi's goals for 2026, and how each quarter landed

Set at the start of the year. Each quarter, a ten-minute check-in records where every goal stands and a band for the quarter.

GoalQ1Q2Q3Q4
Cut first response time on the self-serve queueTarget: Under 4 working hoursAt risk11 hoursOn track6 hoursAhead4 hours in September, holding for six weeks
Write the billing escalation runbook and get it usedTarget: Two other people running them unaided by the end of the yearnot recordednot recordedOff trackDrafted, nobody else running escalations
Bring repeat contacts downTarget: Under 12%not recordedAt risk17%On track13%
Check-in bandMeeting expectationsMeeting expectationsMeeting expectations

The latest check-in (Q3), in the manager's words

Went well
Target hit and holding.
Got in the way
Escalations still all coming to him.
Ravi needs from their manager
The review slot, from me.

Open follow-ups from 1:1s

  • I will book the review slot for the billing runbookmanager · due Sep 5, 2026 · overdue
  • Ravi will pick the two people to train on billing escalationsRavi · due Oct 31, 2026
The record · 1 / 5

Guardrails

Five things the draft may never do.

Checks that run over the finished text every time, whichever writer produced it. Both drafts above pass them.

  1. Quote a colleague

    Any run of eight or more words shared with someone's raw answer is caught — checked against what they actually wrote, not the anonymized copy. With three respondents, nine words identifies the author.

  2. Attribute a view to someone who was not asked

    “Colleagues felt…” is flagged when no colleague input was collected. Your own account of how someone works with others is fine; that is your observation.

  3. Call one person's view a pattern

    One reply is used, and used as one person's view. Two people agreeing is the smallest honest claim of a pattern, and the draft says so.

  4. State a number that is not in the record

    Every figure in the prose is matched against the recorded goals, targets, results and check-ins. This is the check that most often catches something, and a plausible invented number is the most damaging thing a generated document could contain.

  5. Open a formal process

    Warnings, performance plans, dismissal and severance are processes with legal form, and the language for them is rejected outright. An assessment records; it does not initiate.

A draft that fails a check is still shown to you, with the failures listed above it. This is your private preparation, and withholding it would leave you with nothing and no idea why.

Questions

Does the AI write my performance reviews for me?
It writes the prose; you supply everything it is made of. You answer 4 guided questions for a quarter or 7 for a year, in note form, and it turns those notes plus the goals and check-ins you already recorded into a document a person can read. It is given nothing else. It cannot reach a figure you did not record, it cannot invent an occasion, and it cannot set the rating. If you write two sentences you get a document that says it was written from two sentences.
Does it decide the rating?
No, and it is not allowed to try. The prompt tells it to leave the band empty, the provider blanks the field again on the way back in case it ignored that, and the document fails its checks if a band ever appears without a written reason beside it. Below, meeting or exceeding is set by you, in one place, next to a restated definition of what the bands mean — against this person's own goals, not against their colleagues.
Why do my answers come before the colleague input on the page?
Because the order changes what gets written. A manager who reads three colleagues' views and then writes has produced a summary of the colleagues. A manager who writes first has an account of their own, which the colleagues can then confirm or contradict. Only the second is worth having, so the page puts your questions at the top, above the colleague input and above the button that writes anything. The tick-lists come before each text box for the same reason at a smaller scale: read before writing, a list reminds you the quarter also contained an incident and a fortnight of covering for someone.
What does the quarterly assessment ask, and what does the annual one add?
Quarterly: What did they actually deliver this quarter; Where did the quarter fall short of what you both expected, and what happened; What did they get noticeably better at; What did you say you would do for them that you have not done. The annual set keeps those and adds three that only make sense over a year: how they worked with other people, the one change that would make the biggest difference next year, and what next year should look like. 4 of the quarterly questions and 6 of the annual ones carry a tick-list above the box, because a blank box tests recall and recall reaches the last six weeks and the loudest project.
What exactly is sent to the AI, and what is it told?
Generating a draft sends the record to Anthropic's API, and the page says so at the button rather than in a policy page. What travels: your answers as question-and-answer pairs, the goals, targets and quarterly results you recorded, anything you ticked (labeled as your account rather than as evidence), colleague input with the names already stripped, and any summary the person chose to share, marked as their own account of themselves. There is no email address in any of it. What it is told: turn notes into sentences a manager could say out loud, keep colleague input and the person's own summary apart, say plainly what the document cannot support, state both sides of a disagreement and resolve neither, and leave the band empty. All of that is also enforced afterwards by checks that run over the text, because a rule that only exists in a prompt is a hope.
What happens if the model fails or produces something bad?
The output has to parse as the exact document type or it is rejected and retried once with the failure handed back as feedback. If that fails too, a deterministic writer produces the same document from the same inputs with no model involved — the same sections, not a degraded version — and the record says which one wrote it. Then the checks run over whatever came back, and a draft that fails one arrives with the failure named at the top of it rather than quietly. Nothing is hidden from you: you asked for the draft and you are the one who decides what to do with it.
Can I use the manager tools without sending anything to an AI provider?
Not as a per-account switch today — that does not exist yet, and saying otherwise would be wrong. What does exist is the deterministic writer: it produces the whole document from the same inputs with no model involved, and it is what runs when a model is not configured or fails. If sending a written characterization of a named employee to a third party is a problem for your organization, say so at facets.team/contact, because that is the question that would decide whether an option gets built.
Is a tick counted towards anything?
No. No score, no band, no derived total. The counts go into the audit trail so that "where did this come from" has an answer, and are read back by nothing — a checklist in a performance record reads as quantitative to anyone who later opens it in a dispute. Ticking a lot does not buy you a confident document either: the measure of how much the write-up hedges counts prose only, and a test asserts the hedging survives a fully ticked, thinly written intake.
Can I edit what it writes?
All of it, and the document is yours from the moment it appears. You can rewrite any sentence, add sections it did not know to write, delete ones you disagree with, and rewrite the draft from scratch against your updated answers. Today the editor is the document's structured text — every field in one box, blunt but nothing gets lost — rather than a word processor. The checks run again over what you save, and the record notes that the prose was edited by hand, so a document you rewrote is recorded as one you rewrote. Setting the band is not an edit.
Can the person being assessed read it?
You can export the whole document and give it to them, and it is written to be read that way — what it cannot support first, the rationale beside the band, and a closing line saying what fed it. There is no in-app view for them yet. Separately, in most places a written characterization of an employee is personal data they can ask to see, so the honest assumption is that they will read it. The app says so wherever you type about someone.
Is this connected to the team survey or the leadership feedback?
No, and they never join. Those are developmental: people answer under a promise that their individual ratings reach the person they are about and nobody else. This is a performance record. No table on this side references one of those, none of them reference back. The only route between the two is a person choosing to share their own summary with you, which is what the end-of-year example shows.
Where do the examples on this page come from?
The people and every figure about them are invented. Nothing else is: the questions and tick-list labels are read from the same question bank the product uses, the documents are the product's own document type, and they are rendered here by the same component that renders them inside the app. Both are run through the real guardrail checks in our test suite — against the raw, un-anonymized colleague answers, which is the version that could prove a leak — and the suite fails if either stops passing.