The Edu Prompt: AI for the Work of Education
ChatGPT Work, Sites, practical prompts, reusable skills, and new Skill Labs from OpenAI Academy.
Issue 3: July 17, 2026
Welcome back to The Edu Prompt, a field guide for learning, teaching, and building with AI from OpenAI Education.
This week brings new launches, practical workflows, and field notes from education teams around the world.
In This Issue
Feature: ChatGPT Work and what it could mean for education teams.
Product Corner: Sites for turning ideas into simple, shareable pages. Also worth noting from last week: GPT-5.6 and GPT-Live.
Try this with ChatGPT: Make a scene.
Try this with ChatGPT Work: Calibrate an assessment.
Pro-tip: Build a skill.
Around the World: OpenAI Education for Countries meets leaders in Italy.
Learning with OpenAI: New hands-on Skill Labs.
Feature Story: Introducing ChatGPT Work
Faculty and staff around the world already use ChatGPT to brainstorm, draft, analyze, and solve problems. ChatGPT Work, powered by our latest frontier model, GPT-5.6, extends those familiar workflows into projects that involve multiple steps, sources, tools, and deliverables.
AI is becoming less about one-off answers and more about helping people organize, draft, analyze, and reuse work across the flow of a day. For educators, staff, and leaders, that matters because the bottleneck is rarely a lack of ideas. It is turning those ideas into something usable: a lesson plan, a family message, a grant outline, a meeting brief, a rubric, a project plan, or a first draft that can be improved.
ChatGPT Work is designed for tasks that need context, tools, and follow-through.
Product Corner: Build a webpage with Sites
With ChatGPT Work, Codex, and Sites, you can move from a rough idea to a lightweight webpage that is easy to share. For education teams, that can mean a resource hub, project tracker, event page, course companion, or internal guide without starting from a blank page.
Example
Create an experience focused on W. B. Yeats’s poem “The Song of Wandering Aengus.”
Include the full public-domain poem, a brief Yeats biography, historically appropriate
images, and interactive annotations on selected lines and phrases.
Style it with an early-20th-century Arts and Crafts / Celtic Revival look,
using aged paper textures, restrained ornament, period typography, and accessible contrast.
Build the site directly in @sites and return an editable result.
Try this with ChatGPT
Create a time-travel classroom: Use ChatGPT to generate images that help students bring ideas, people, and concepts together. Use this as a discussion starter, not a historical source.
Create a realistic classroom where Leonardo da Vinci, Ada Lovelace,
Albert Einstein, Maya Angelou, and Marie Curie are all students learning together.
Educational, respectful, historically recognizable.Fine-tune output: Once you get an image, tap “Edit” and use the comment feature to select areas you’d like to modify.
Try this with ChatGPT Work
Calibrate assessments: Review scoring patterns without assigning or changing grades.
Review the de-identified submissions, rubric, grading export,
sample feedback, course outcomes, and calibration rules I provide.
Create a calibration workbook that:
- flags possible scoring inconsistencies
- summarizes common strengths and gaps
- identifies feedback themes
- suggests rubric clarifications
- cites evidence for every flag
Do not assign or change grades. Route every scoring judgment to faculty review.Don’t forget to validate the findings: Faculty reviewers should check the source evidence, confirm that comparisons are like-for-like, and look for missing context such as accommodations or assignment variants.
Open the calibration workbook prompt in ChatGPT
Pro-tip: Build a skill for repeated workflows
You can turn a repeated workflow into a reusable Skill that you can invoke later by name. Here’s an example for prioritizing your inbox.
Help me create a reusable skill that reviews recent email and prioritizes what needs my attention.
Before creating it, ask me one question at a time about:
- Which inbox and date range to review
- What counts as actionable or urgent
- People or topics to prioritize
- The output format I want
- What actions the skill may take
Suggest sensible defaults when I’m unsure.
Then summarize the proposed workflow for my approval before creating the reusable skill.
Around the World: Field Notes from Italy
Last week, OpenAI’s Education for Countries team traveled across Italy to learn from and strategize with university leaders, professors, PhD students, and local businesses to imagine and build the future of education with AI.
Each stop—Turin, Modena, and Naples—revealed the unique culture, industry, educational focus, and aspirations of each region. AI is most powerful when it amplifies the experiences, skills, and intentions of people and communities.
Learning with OpenAI: New Skill Labs
Join us in OpenAI Academy’s newest learning series, Skill Labs, where you’ll get hands-on practice with everyday workflows you can use immediately, whether you’re faculty, staff, working in student support, or an administrator.
07.30 Automating workflows
08.06 Analyzing data
08.11 Managing daily work
Have a story idea?
We are looking for practical examples from educators, faculty, staff, students, and education teams using OpenAI tools to save time, improve learning, or build something new. Reply with what you tried, what changed, and what another educator or team could reuse.
Thanks for reading. If this issue gave you something useful to try, share it with an educator, staff member, or team who might use it too.
The assessment calibration example is the part I keep coming back to.
AI doesn’t need to replace the grader to be useful. Even surfacing inconsistent scoring, recurring misconceptions, and unclear rubric language could make assessment much fairer.
The important boundary is exactly the one you’ve kept here: let the system reveal patterns, but keep the final judgment accountable to a human.
That feels like a much healthier direction for educational AI—less “let the model decide,” more “help us see what we were missing.”
This summer, working with Dr. Amine Amar, we have been testing Fikira (https://framethechangelab.com/fikira-download.html) with three faculty members in Morocco, funded by Al Akhawayn University, and because two of them teach in French and one in Arabic we have had to think from the start about a tool that cannot quietly assume everyone works in English.
Fikira is a free learning companion that runs on the teacher's own computer and keeps working when the internet drops, which matters because the teachers and students we most want to reach do not have reliable broadband or the budget for an AI subscription. A teacher gives it a syllabus, and it looks for the assignments a student could now finish with AI without doing any of the thinking the assignment was meant to develop, then suggests ways those assignments might be redesigned. We kept students out of this first phase on purpose, since it seemed wrong to put the tool in front of a class before the people who actually teach the course had told us whether it was worth their time.
The reality has been harder than the idea, mostly in ways we did not see coming. A local model has a training cutoff and has never seen the course in question, so its early advice was fluent, plausible, and could have been written about almost any syllabus anywhere. Then in June we hit something considerably more embarrassing, when the analysis timed out on a teacher's laptop and a fallback message we had written for our own debugging was printed into a generated syllabus in the course description field, where students would have read a sentence explaining that Fikira had used a fast local analysis because the full one was unavailable. Around the same time we found that the tool had lifted a sentence about grading percentages out of a syllabus and was analysing it earnestly as though it were an assignment.
What helped in the end was not the larger model we assumed we needed, but taking away the model's licence to assert anything it could not point at. Fikira now pulls the assignments out of the document before the model is involved at all, quotes the wording behind every finding it reports, and says plainly that something is not evidenced when there is nothing in the file to quote, so that when it cannot tell whether an exam is supervised or taken at home it says so rather than quietly deciding for itself.
If we could pass on one thing to anyone building something similar, it would be to test on a real teacher's syllabus rather than the tidy examples you wrote yourself, because ours passed for weeks while an actual course document was failing in ways we never saw.
We are still in the middle of this. The teacher survey went out at the end of June and we are waiting on responses, and what those teachers tell us will decide the shape of the next phase, which is a pilot with their students once the faculty are satisfied the tool is ready for them.