Right now · 75 minutes · You are the author
Leave with a small activity for your own course that you have built, tested, and saved.
- 10:10 · 5 min: how the AI is guided, and what an activity looks like underneath.
- 10:15 · 15 min: watch an activity being built and revised with an LLM, with a tour of the tools.
- 10:30 · 45 min: your turn. Plan, build, test in the Sandbox, and save.
- 11:15 · 10 min: three groups share one question each.
Deliverable: a 2–3 question activity whose feedback prompts behave the way you intended, with its source saved somewhere you own.
Four Layers of AI Guidance
Every AI response a student sees is shaped by four layers. Layers 0–2 set the context. Layer 3 is what you write for each question.
-
0
Platform Guidance
Always on, set by coLearn-AI. Core rules: nudge, never solve; match the students’ language; one focused question at a time.
-
1
Class Guidance
Set once per course by the instructor: the teaching contract for that class.
-
2
Activity Guidance
Set per activity: topic scope, student level, and how strict to be.
-
3
Question Guidance (your main job)
Set per question: a feedback prompt and a sample response, both hidden from students. Together they decide whether a submission is accepted or the group is asked to revise.
Key insight: You don’t write the AI’s replies. You write the criteria it uses to decide how to reply.
What an Activity Looks Like Underneath
Every activity is a plain-text document in a LaTeX-like markup. You rarely need to write it yourself, since the AI Builder, an LLM and the Visual Editor all produce it for you, but it helps to recognise the shape. Here is one complete question:
\questiongroup{What does range() produce?}
\question{Before running it, predict exactly what this prints. Then run it and explain any difference.}
\python
for i in range(2, 10, 3):
print(i)
\endpython
\textresponse{4}
\sampleresponses{It prints 2, 5, 8. 10 is the stop value and is never included.}
\feedbackprompt{Accept if the group gives 2, 5, 8 and says why 10 is not printed. If they include 10 or 11, ask what the stop value means.}
\questiontype{output_prediction}
\endquestion
\endquestiongroup
Students see the question, the code and an answer box. They never see the sample response or the feedback prompt: those are for the AI. The optional \questiontype line is for research: it labels what kind of thinking the question asks for (here, predicting output), so the trace can be analyzed by question type. It changes nothing for students or for the AI. See Research.
Build It, Then Revise It with an LLM
Watch an activity being created from scratch and improved in a round or two, using the tools described below. You’ll do the same thing in your group.
- Describe it to the AI Builder. Topic, students, how many questions, what reasoning they should show. The platform generates a first draft.
- Look at it in the Visual Editor. What’s good, what’s missing, which feedback prompts are too vague?
- Click Copy for LLM. This copies the activity plus a briefing on the markup. Paste it into ChatGPT, Claude or any LLM and ask for a specific change: “add a prediction step before each code question,” “tighten the feedback prompts,” “add timings.”
- Paste the reply back and click Review Pasted Revision. The platform lists what the revision adds and removes. Accept it only if nothing important was lost.
- Repeat a couple of times. Each round is a conversation you can see and push back on.
Your instructions go to the LLM, not into the paste box.
The paste box takes the LLM’s revised activity. Type what you want changed in the LLM, alongside what Copy for LLM gave you.
The Authoring Tools
👁 Visual Editor
The activity as students see it. Click any question or AI block to open it in the Question Panel, edit its fields, or have the AI revise just that question. Try Changes applies your edits; Save keeps them.
✏️ Source
The markup itself. Useful for small, precise edits; errors are reported by line. You never have to use it.
▶️ Sandbox
Runs the activity right inside the editor, as a student, with real AI feedback. This is how you find out whether your feedback prompts work.
Six Ways to Create and Edit an Activity
📖 Reference: not covered live. Read this after the workshop, or skim it if you finish early.
coLearn-AI activities are plain-text markup files. That means you can create and edit them through several different workflows. Choose the one that fits your style and how much of the markup you want to write by hand.
Option 1
Internal AI Creation
Use the AI Builder in the creator. Describe your topic, students and question goals; the platform generates a draft activity you can open in the Visual Editor. Best for getting started quickly.
When to use: First draft, unfamiliar with markup syntax.
Limitation: Generated prompts often need hand-tuning for Layer 3 precision.
Option 2
External AI with Example + Syntax
Paste the markup language reference and a complete sample activity into ChatGPT or Claude, then ask it to write an activity for your topic. External models have broad domain knowledge and generate coherent question sequences quickly.
When to use: You know your topic well but want a full first draft fast.
Tip: Tell the AI your student level, topic constraints, and which question types to use.
Prompt template: “Here is the coLearn-AI markup syntax [paste reference]. Here is a complete example activity [paste source]. Write a 3-question activity about [topic] for [student level]. Include feedbackprompt and sampleresponses for each question.”
Option 3
Visual Editor
The Visual Editor shows the activity as students will see it. Click a question or AI block to open it in the Question Panel, edit its fields or ask the AI to revise just that question, then Try Changes and Save. No markup required.
When to use: Fixing a specific question, renaming a section, adjusting Layer 3 guidance on an existing activity.
Limitation: Building a new activity from scratch is faster with Option 1 or 2, then refine here.
Still under development: an edit can occasionally garble the activity or produce an error. If that happens, press Undo (the curved-arrow button) to step back to the last good version.
Option 4
Copy for LLM → Paste Back
Click Copy for LLM to copy the activity together with a syntax briefing. Paste it into ChatGPT, Claude or any LLM, describe the change you want, and keep going until you like it. Paste the reply back: the platform shows what the revision adds and removes before anything is saved.
When to use: Reorganizing question order, converting question types, or regenerating feedback prompts across multiple questions at once.
Tip: Use Duplicate First to try a big rewrite on a copy, leaving the version students use untouched.
Option 5
Direct Source Editing
Open the Source view and edit the markup by hand. Full control over every tag, with line-level error messages. See the Markup Reference.
When to use: Experienced authors, fine-tuning edge cases, or when you know exactly which line to change.
Tip: The markup is intentionally readable. After a few activities, direct editing is often the fastest approach.
Option 6
The Sandbox
The Sandbox runs your activity right inside the creator, as a student would see it, with real AI feedback. Submit test answers and see whether the AI accepts them or asks for a revision, before any student does.
When to use: Calibrating a feedback prompt; checking edge cases; seeing the student experience.
Tip: Test three answers: clearly correct, plausible-but-wrong, and off-base.
Recommended workflow: Draft with the AI Builder or an LLM → revise with an LLM (Copy for LLM) → polish in the Visual Editor → test in the Sandbox → fine-tune in Source if needed.
Create, Test and Save Your Own Activity
Work with your table, or on your own. Aim for something small that works: 2–3 questions on a topic you teach, tested in the Sandbox. Three good questions beat ten untested ones.
Split the work at your table
- Driver: builds and edits the table’s activity on their own creator account.
- Planner: keeps the Step 1 plan, and checks each feedback prompt against it.
- Tester: writes the three test answers per question for Step 3, and reads the AI’s replies aloud.
- Timekeeper: calls each step on time. Saving at 11:10 is not optional.
Step 0: Switch Accounts (10:30, 3 min)
Everyone logs out of the student account and logs in at colearn-ai.com with the creator login on their card. Then open the CCSC:MW Create class (code: create). It starts empty, and it is where you create and test your activity. Each person has their own creator account, so after the workshop anyone can build an activity of their own.
A browser holds only one login at a time, so logging in as your creator replaces your student login in every tab. To keep both open, use a second browser or a private / incognito window.
Step 1: Pick a Topic and Plan (10:33, 7 min)
Choose one idea your students find hard, in any subject you teach (examples from other fields). Before you touch the tool, decide for each question:
What should students do? predict / explain / write code / compare / critique
What does a good answer contain? this becomes the sample response
What is the one thing that separates good from not-yet? this becomes the feedback prompt
What common mistake should the AI ask about? e.g. “includes the stop value”
Step 2: Build a Draft (10:40, 10 min)
- Create a new activity in the CCSC:MW Create class.
- Describe it to the AI Builder using your plan from Step 1. Or use Copy for LLM with your own LLM.
- Need an example of a feature? The coLearn-AI Feature Gallery in the CCSC:MW class (with your student login) shows every feature, numbered by type. Its source is also on the Resources page.
- Check every question in the Visual Editor. Does each one have a sample response and a feedback prompt you agree with?
- The Visual Editor is still under development. If an edit messes something up or shows an error, press Undo and try again, or make that change with Copy for LLM instead.
Step 3: Test in the Sandbox and Revise (10:50, 20 min)
This is the most important step. Open the Sandbox and submit three answers to each question:
- A clearly correct answer. It should be accepted.
- A plausible but wrong answer. It should be sent back with a question that points at the mistake without giving it away.
- An answer that misses the point. It should also be sent back.
If the AI gets it wrong, tighten the feedback prompt, either in the Question Panel or with another LLM round, and test again. See Writing a Good Feedback Prompt below.
Step 4: Save Your Work (11:10, 5 min)
The workshop accounts are temporary, so take your activity with you.
Open the Source view, select all, copy, and paste it into an email to yourself or into a document. That text is the activity. Later, paste it into the Source view of a new activity in any coLearn-AI account.
Writing a Good Feedback Prompt
Feedback prompt (hidden from students)
“Accept if the loop uses range() to produce 3, 6, 9, 12, 15, 18. If the stop value leaves out 18, ask what range() does with its stop value.”
Sample response (hidden from students)
for value in range(3, 19, 3):
print(value, end=" ")
One criterion
Name the single thing that separates a good answer. If the prompt needs “and” several times, split the question.
Ask, don’t tell
Tell the AI what to ask about when something is missing: “ask what the stop value means,” not “explain the stop value.”
Show what’s equivalent
If there are two genuinely different correct answers, give both as sample responses, or say what to accept.
Design for the trace
Decide what reasoning you want to see in the epistemic trace (a prediction before running code, say) and ask for it.
Share with the Room
Three groups share, about 3 minutes each. Each group shows one question: what students see, the feedback prompt behind it, and how the AI responded in the Sandbox. Was that what you expected, or a surprise?