Earth Day × Six Thinking Hats

For Earth Day week, I wanted seventh graders to discuss Earth Day Taiwan's "Million Green Actions" in groups and commit to one action they would actually do. Ordinary group discussion tends to be carried by a few voices while others drift. Classroom engagement tools such as Mentimeter, Kahoot and Pear Deck are good at quizzes and polls, but not at walking a group through a full line of reasoning.

So instead of a quiz, I built the Six Thinking Hats framework into the flow itself: white (facts) → red (gut feeling) → yellow (value) → black (risk) → green (ideas) → blue (decision). That takes students from evidence to a concrete plan, step by step.

• Students: seventh graders in three classes (704, 708, 717), about 87 in total, answering on tablets in small groups.
• Teacher (me): needed to see each group's progress in class, and afterwards to edit, review and export results and assess the quality of thinking.
• Setting: real lessons during Earth Day week, with limited time, uneven devices and network, and a wide spread of ability. Six actions from the official list of 42, grouped as plastic reduction, carbon reduction and saving resources.

• No engineering team: I owned the instructional design and requirements, and the code was built with Claude. The stack had to stay simple: plain HTML/JS plus Firebase.
• Used while being built: fixes shipped between and during lessons, and nothing could break a class.
• Student data: the console can edit and delete answers, so it needed real authentication. Nothing public may identify a student.
• Cost: AI analysis runs after class in batches on Azure OpenAI, not live.

• Sequential unlocking: the next hat opens only after the right teammates save, so nobody skips ahead.
• Group-locked actions: everyone in a group discusses the same shortlisted actions.
• Assigned roles → open participation with Pass: at first each student owned one hat. In class, students without a hat tuned out, so I switched to whole-group answering with an explicit Pass.
• Green hat sees black hat: the green prompt asks for fixes to black-hat concerns, so those concerns now appear live on the green screen.
• Blue-hat consensus: the final plan needs consensus, but only the students who disagree re-choose.
• Firebase Authentication replaced a password hard-coded in the console.
• Coherence over AI detection: I assess whether answers connect across hats, not whether AI wrote them.

I defined problems and rules, and Claude implemented and debugged. Three patterns:

• I observe, it locates: after the switch to open participation, Save got locked. I reported the symptom, and Claude traced it to subscribeHatReadiness and a hatKey race condition.
• I set the teaching goal, it builds the UI: I wanted green to answer black, so Claude added a live black-hat reference panel.
• I define the frame, it builds the tool: for post-class analysis I set three questions (yellow cites white, green answers black, blue is concrete). Claude wrote a script that computes the stats in code and then hands them, with the raw answers, to Azure OpenAI for interpretation.

• Used in real Earth Day lessons across 3 classes, about 87 students.
• Class 708 analysis (28 responses):
  • Full-completion rate by hat: white 36%, red 57%, yellow 46%, black 50%, green 54%.
  • Final choices: Action 1 (carry your own bag, utensils and bottle) with 13 students and Action 32 (take the stairs) with 10.
  • 6 of 8 groups reached blue-hat consensus.
  • Three teaching recommendations: a white-hat data template, a blue-hat plan template and peer review.
• The framework doesn't depend on the topic. It can carry SWOT, literary critique or civic-issue debates.

• Scaffold white hat from day one. At 36% it had the lowest completion. A data-and-source template should come first, not after the report.
• Give blue hat a format covering what, when and how to check, because some plans read like slogans.
• Compute more in code. In the report, the model's own count of groups by coherence level didn't add up. Every number should come from code, and the model should only interpret.
• Handle secrets from the first API call. The analysis script first parsed its key from a local document and only later moved to environment variables.