OpenAI engineer uses Codex to automate his own performance review

Headline

OpenAI engineer Jason Liu shows Codex automating his performance review prep using Slack and Gmail.

Summary

Jason Liu, who recently joined OpenAI as a Staff ML Engineer, posted a tweet showing Codex handling a genuinely tedious task: putting together a performance review promo package. Liu created Instructor, a Python library for structured LLM outputs that gets over 6 million downloads a month and influenced OpenAI’s own structured output features. In the demo, he had Codex pull information from Slack and Gmail to assemble the document. It’s a small example, but it shows where these tools are headed—handling the boring multi-step administrative work that eats up engineering time.

Analysis

Liu isn’t just any engineer posting about AI tools. He built Instructor, which became popular enough that OpenAI incorporated similar ideas into their structured outputs. Now he’s inside the company, and this tweet reads like someone using their own tools for actual work rather than a polished demo.

The interesting part is the multi-tool aspect. Codex isn’t just generating code here—it’s coordinating across Slack and Gmail, pulling together information and assembling something useful. That’s closer to how people actually work: jumping between tools, gathering context, synthesizing it into something coherent. OpenAI has been pushing plugins and background automations, and this fits that direction.

There’s a personal angle too. Liu has talked about using AI to get back into technical work after a hand injury. So when he demos something like this, there’s a “I actually need this” quality to it.

The obvious questions: how well does this work when the task is messier? What happens with sensitive data flowing through these integrations? But as a proof of concept for AI handling administrative grunt work, it’s concrete.

Impact Assessment

• Significance: Medium
• Categories: Developer Tools, Technical Insight, Industry Trend