At my company we give ourselves peer, upward, and self performance feedback.

I don’t mind this practice because I understand the value that it can bring in terms of career development and team cohesion. But writing reviews can take quite a bit of time, especially for more senior members of the team. We do it twice a year in the spring and fall.

This cycle, I looked into using generative AI assistants to help me write feedback. So far this process is much preferable to writing it without LLMs.

Some thoughts and tips…

  • Provide data sources to the AI about what a colleague worked on
    • AI assistants can process large volumes of even unstructured data very quickly
    • Data exists on teammates’ contributions, though some might be easier to access
      • Tickets created and closed by individuals
      • Daily standup notes (we use Geekbot for this)
      • Slack discussions
      • Docs, sheets, wikis, other artifacts
      • Code and code reviews
      • Prior review feedback
    • Data helps ground the feedback and make it more specific
    • Also helps you to recall what somebody worked on
  • Prompt with relevant information
    • Like the colleague’s role/level, your role and relationship to them, how long they’ve worked on the team, their main contributions, etc
    • Your company’s expectations around performance in different capabilities at different levels (career matrix)
    • Your team’s goals
    • Your company’s examples of well-formed performance feedback
  • Provide subjective thoughts on the colleague
    • Recall how it felt working with them
    • What you thought some of their core strengths and weaknesses have been
    • I use voice mode with a stream of consciousness style discussion here to generate ideas and let the AI clean it up later
    • Ask the AI to challenge your beliefs against the data you have (eg does Sally really get bogged down in details or does her work history show she has a good balance with strategic planning?)
    • Work with the AI to uncover nuances in teammates’ performance that might be valuable to recognize
  • Create reusable tools for yourself and others
    • A Custom GPT in ChatGPT, links to collect data, or useful prompts can be shared

I do think this approach risks turning into that meme where like I ask my chatbot to translate my bullets into a paragraph and my manager asks their chatbot to translate the paragraph back into bullets. But I’m trying to avoid that outcome by tuning the output to be concise.

So far I feel like this process has helped me write better performance feedback more effectively. Maybe I can return to the same chats in the next cycle or mid cycle to track performance over time.