What they did
Cyrus Wong, a senior lecturer at HKIIT who writes dozens of transfer recommendation letters a year, built an agent called AutoFolio on Gemini's computer-use model through Vertex AI. Instead of logging in itself, it attaches with Playwright to a Chrome window where the lecturer is already signed in to LinkedIn. It then works through the profile in five stages (header, honors, volunteering, certifications, posts), clicks to expand hidden text, takes screenshots as proof and downloads event photos. From that evidence it writes a recommendation letter, an HTML and Markdown portfolio, a supporting document and a JSON manifest. A batch mode reads a list of student profiles and builds a class dashboard.
How it went
Wong ran it on one student, Hiu Tung Cheung, who was applying to HKUST computer science and agreed to be the test case. The agent found the student's community leadership roles, a PyCon Hong Kong talk, scholarships and grade average, competition prizes and 16 cloud certifications, and produced a 1,200-word letter tied to those items. The outputs are published in the repo. The post does not say whether the letter was sent as written or how much was edited, and gives no run time or cost.
Worth knowing
Attaching to your own logged-in browser avoids 2FA and stored passwords, but the agent then acts with your full account. Wong turns on the SDK's prompt-injection detection because profile text is untrusted. Only run it on people who have agreed, and check every claim before signing a letter.