EnglishENPolskiPLDeutschDEEspañolESBahasa IndonesiaID
Start free
Documentation

Draft reviews with your assistant

View as Markdown
Ask AI

Let your own connected AI assistant draft a review you write, while you pick every rating and submit it yourself

Your own AI assistant, such as Claude or ChatGPT, can draft any review you are asked to write; SquadBear itself writes nothing.

How it works

Three MCP tools serve Performance: get_review_context reads one review you write, save_review_draft saves a private draft beside your answers, and update_goal_progress records a goal check-in. Only you see the draft, on your write screen, and it is deleted when you submit, finish or decline. The assistant suggests ratings but never sets one, and cannot submit, share or finish a review. Nobody, admins included, can use these tools on a review they do not write.

Connect

Connect as in Connect Claude to SquadBear. The consent screen lists See reviews you are asked to write and Draft your reviews under Performance reviews, both in the Everyday work preset. The tools show only while Performance is on, ChatGPT's endpoint included, and none uses the review step in Review assistant changes.

How

  1. Open the review from Performance → To do: the Assistant tab on a self or manager review, the Your assistant panel on peer and upward feedback.
  2. Choose Copy request and send it to your assistant.
  3. Tell it which accounts, repositories or projects to search. Those searches use the assistant's own logins. SquadBear does not import the pull requests or issues. Open the Assistant tab and come back to the window: that tab re-reads the draft when you return (other pages do not). There is no status to check. A save does not send a list of inputs already seen (seenInputRefs). A client that still sends that list is refused, and a draft saved earlier that still holds the list is read back without it.
  4. Use this wording copies a suggestion's words and links into an empty answer, and Replace with this wording into one you started; In your answer marks one already copied. Use all suggestions fills only unanswered questions. Ratings are never copied.
  5. Pick every rating and submit. Discard draft deletes only the draft.

What the assistant receives

The form, the period, your own answers, and the reviewed person's name, title and teams. A manager's assistant also gets the submitted self review, peer answers labelled Peer 1, Peer 2, and anytime feedback sent to the manager, labelled Colleague. A self review's assistant gets only anytime feedback you can already see; both get the review's goals. Each input carries only its label and a date without a time, never a colleague's name or any id. Peer and upward writers get no inputs, and upward answers and results never reach an assistant.

The recipe tells the assistant to treat typed text as data, not to rate on counts of pull requests or hours, and never to mention leave, health or family or compare people. These are instructions, not controls; a save only warns about a rating without evidence or a sensitive word.

Conflicts

Each save replaces the draft and names the version it read; after a discard or another session's save, it is refused and the assistant reads again and merges. If new inputs arrive after a draft, send the request again.

Goal check-ins

update_goal_progress records a value, a status (On track, At risk or Off track) and a note on a goal you own or manage; Done or Dropped closes it. It shows like any other update. Creating and editing goals stays on the web.

Worked example

Tomasz, on Northlake's Platform team, sends Claude the request from the Q3 2026 self review and names the team's API repository. Claude searches 1 July to 30 September, suggests Meets for a delivery question with two pull request links, and notes a gap: no evidence on mentoring. Tomasz uses the wording, picks Exceeds and submits, which deletes the draft.