From Custom Gems to Autonomous Agents: Upgrading My AI Workflow Across Classroom, Campus, and Research

 

用學生優惠升級Gemini PRO的這一個禮拜以來，每天很有興趣地在嘗試錯誤中學習，摸索Spark的各項主動代理的功能。之前就很喜歡自建Gems的流程，用同樣的觀念在今天成功地試出Tasks真的可以自己一步步完成工作，超級開心! 接下來，想要根據自己不同的學習與工作角色，改用Skills來自動化完成。

Are we using artificial intelligence to reclaim our time—or have we just signed up for a second job babysitting our screens?

For over a year, educators have been told that the future belongs to those who master prompt engineering. But what if the real breakthrough in educational technology isn’t learning how to write better prompts, but learning when to fire yourself as the prompt engineer?

Midnight in Taichung: The Digital Swamp

It is 11:45 PM on a Sunday in Taichung. You are sitting at your desk with cold oolong tea, staring at a Google Drive holding twenty years of your teaching life—CAP-style reading passages, bilingual biology cards, Taichung English Advisory Team workshop slides, and National Chung Hsing University doctoral seminar PDFs.

Every seasoned teacher knows that sinking feeling: you have forty minutes before bed to prep tomorrow's class, and you blow twenty hunting for an unfiled handout from 2022.

The First Test: Handing Over the Keys

When I used my doctoral student status to upgrade my Gemini workspace to Pro, my first experiment wasn't writing an academic paper. I challenged Gemini Spark to crawl twenty years of teaching files, tag each file's pedagogical target and CEFR level, sort them across my professional roles, and rebuild the drive into a clean directory tree.

Watching the agent execute that overhaul quietly in the background forced a sudden realization: the bottleneck in my workflow was never the software. It was me.

The Gem Pipeline: Building Multi-Step Assistants

Before adopting autonomous agents, I built custom Gemini Gems for lesson planning and doctoral assignments through a four-stage workflow:

1.    Visualize the terminal artifact. Start with Backward Design. Clarify the finished reading task or rubric before drafting instructions.

2.    Spar through dialogue. Refine constraints and tone through active back-and-forth testing.

3.    Chain multi-step stages. Build pipelines where the model digests raw text, calibrates vocabulary to CEFR B1, and formats comprehension tasks.

4.    Stress-test with messy inputs. Test edge cases until outputs reliably match classroom reality.

I shared these Gems at workshops across central Taiwan for Teaching English in English (TETE), CAP reading sets, and teacher recruitment interviews.

Shifting to 24/7 Background Agents

Yet even custom Gems wait for you to press Enter. Gemini Spark runs asynchronous background workflows across Drive, Docs, and Calendar:

1.    Academic monitoring. Tracking cultural study preprints and filing reading digests into my dissertation folder.

2.    Advisory triage. Parsing teacher workshop feedback across the city to draft briefing memos.

3.    Curriculum prep. Drafting differentiated vocabulary glossaries the moment a unit appears on my calendar.

Balancing a high school classroom, an advisory team, university lecturing, and doctoral research means time is my scarcest asset. Automating administrative routines is an act of pedagogical triage. Our energy belongs where learning happens: coaching hesitant speakers and guiding teachers through real hurdles.

Let background routines clear the underbrush so you can do the actual work. HERE WE GO!