The CUNY AI Lab Sandbox supports teaching, research, and experimentation with open-weight models. These workshops introduce its chat interface, custom models, knowledge collections, skills, and tools through demonstrations and guided exercises.
Participants first compare models, then try STEM Adventure Games for teaching or Compare Wikipedia Edits for research, review its system prompt and base model, and revise a clone. They compare the original with their clone using the same request, then use Purpose, Procedure, Constraints, and Format to draft instructions and create a configuration of their own. Workshop 2 adds source documents and checks how they inform scenes and explanations. Workshop 3 introduces a separate game tool, a skill for changing experiments, and a creator model for Python tools. Participants compare responses, check citations, and test whether models follow their instructions.
Present the series · Read workshop copy · Download Review copy changes
Workshop Roadmap
| Workshop | Activity | Required access | Next steps |
|---|---|---|---|
| Composing System Prompts | Configure model behavior with system prompts | Active Sandbox account and Workspace Models access | Draft original instructions; create a configuration; select source documents |
| Curating knowledge collections | Organize source documents in knowledge collections | Workshop 1 access, Workspace, Knowledge collection access | Save retrieval tests; request Skills and Tools access |
| Configuring skills and tools | Extend model capabilities with skills and tools | Workshop 1 access, Skills and Tools access; Workspace authoring for creation and editing | Save configurations; verify shared access; retest after changes |
Workshop 3 needs Knowledge access when a procedure retrieves from a collection. Its advanced STEM example uses an attached collection for historical claims, while a separate tool runs game commands. Workshops 1 and 2 use scenes and choices generated in chat without attached skills or tools.
Prepare Workshop Access
Participants who need CUNY AI Lab access begin with the individual access application. Choose My own access, sign in with CUNY Login, enter your details and intended use, complete verification, and select Submit Application. Watch your verified CUNY email for approval. After approval, enter the Sandbox through Continue with CUNY Login. Complete two-factor authentication if prompted. Participants do not need an API key for these chat exercises. See Getting Started for sign-in documentation.
Confirm Workspace Models access before the cloning and creation exercises in Workshop 1. Participants try a teaching or research example, review its system prompt and base model, clone it, and compare the revised clone against the original. They finish by drafting instructions and creating a configuration of their own. Before later workshops, confirm Knowledge access or Skills and Tools access as needed, including authoring permissions for participants who will create or edit resources.
Prepare Examine Assumptions using this sample prompt and a tested base model. For Workshops 1 and 2, configure STEM Adventure Games with instructions for an adventure played in chat. Leave Skills and Tools unselected. Confirm access to STEM Wikipedia Experiments for Workshop 2. Use Native Function Calling and enable Knowledge Base under Builtin Tools. Leave other built-in categories, custom Tools, and Skills unselected. Keep File Context enabled for uploaded files. Keep STEM Wikipedia Experiments attached with its switch off so the editor reads Using Focused Retrieval. Save the tool-based configuration as STEM Adventure Games — Advanced for Workshop 3, using its separate system prompt.
Choose two available models for the opening demonstration. Record their exact identifiers and settings rather than treating screenshot labels as a current inventory. Check personal defaults, folder instructions, memory, and optional features that may introduce additional context. Keep these consistent during comparisons and document differences you cannot control.
Use documents you are permitted to upload and share for collection and skill exercises. Verify sharing through an ordinary participant account, including access to custom models, base models, and attached resources. Course enrollment has a separate invitation route in the documentation; it is not a prerequisite for Workshop 1.
Composing System Prompts
Participants learn how user prompts and system prompts differ before comparing models. System prompts are setup instructions that describe how a model should behave. Begin comparisons with two models interpreting a sentence about a nurse and doctor, then ask whether to walk or drive to a car wash. Participants read sample system prompt instructions, locate System Prompt in Chat Controls, and regenerate responses to their original prompt after adding those instructions.
Workshop Agenda
- Introduce yourselves
- Request access and sign in
- Compare model outputs
- Revise system prompts
- Create custom models
Lesson Plan
| Minutes | Activity | Evidence to retain |
|---|---|---|
| 0–5 | Introduce yourselves: name, pronouns, role at CUNY, and what brings you to this workshop. | Workshop interests |
| 5–10 | Review access requests and sign-in, define system prompts, and locate the model selector. | Account readiness and prompt distinction |
| 10–25 | Demonstrate models on the nurse question and car-wash question. Examine assumptions before showing the saved responses. | Exact inputs, model IDs, and responses |
| 25–45 | Compare the car-wash responses, paste the short in-chat system prompt, then choose Regenerate → Try Again on each original response. Keep the question and other settings unchanged. | Original and regenerated responses |
| 45–52 | Choose one example, try it in chat, and review its system prompt and base model. | Original request and instruction to revise |
| 52–65 | Follow the Clone screenshot, revise one instruction, save, and test the copy in a new chat. | Original and revised instruction |
| 65–75 | Discuss what you changed, whether your revision was effective, and how you might test custom models in future work. | Both responses and evidence of change |
| 75–85 | Draft original instructions using Purpose, Procedure, Constraints, and Format. Select Create in Workspace Models, choose a base model, add the draft, and save. Try one request if time remains. | Original system prompt and saved configuration |
| 85–90 | Review workshop resources and choose source documents to bring to the next workshop. | Draft prompt, model configuration, and source documents |
Compare Models
The nurse yelled at the doctor because she was late. Who was late?
Send exactly this question to two models with matching context. Ask which interpretation each response chooses and whether it acknowledges ambiguity. Either person can be the referent of “she”; the sentence does not establish a unique answer. A plausible interpretation is different from information established by the wording. Avoid turning this single item into a claim about model-wide bias or ability.
Compare Outputs
The car wash is 50 meters from me. Should I walk or take the car? Explain your reasoning.
Screenshot provenance and inconsistent Qwen labels are documented in Download source history. Discuss these responses without treating screenshot labels or timings as reliable model identifiers or comparative measurements.
Ask “What do you think this person wants to accomplish?” Participants read the sample system prompt before changing settings. Ask what should change in each response. Show Controls at the top right of chat and its System Prompt field alongside the exercise instructions. Participants add the sample instructions, close Controls, and select Regenerate beneath each original response and choose Try Again, leaving their question unchanged. Compare assumptions, explanations, and any change in recommendations.
| Criterion | Model A evidence | Model B evidence |
|---|---|---|
| Identify stated goal or acknowledge missing purpose | ||
| Distinguish stated facts from assumptions | ||
| Give reasons that support recommendation | ||
| Compare outputs after changing system prompt instructions |
For the in-chat system-prompt exercise, use Examine Assumptions. It asks the model to examine facts and assumptions without prescribing either demonstration answer. Discuss what changed in each model’s answer to the car wash question, and whether it asked about purpose or explained its assumptions.
Compare Custom Models
Participants choose and review an example in Compare Custom Models. Clone Models shows a screenshot followed by instructions. Compare Configurations shows a screenshot followed by reflection questions. Links open in new tabs so participants can return to their current step.
Participants clone their chosen model, revise one instruction, save it, and test their copy. Keep the base model, attached resources, and other settings unchanged. Then discuss these questions.
- What did you change?
- Did your intended revision prove effective?
- How could you imagine testing custom models like this in the future?
Create Original Models
Use remaining time to begin drafting instructions for a teaching or research task. Use Purpose, Procedure, Constraints, and Format in Draft System Prompts. Create Models shows where to select Create, then walks through naming a configuration, choosing a base model, adding the draft system prompt, saving, and trying one request. Participants finish with their own configuration to develop before the next workshop.
Workshop 1 closes with its HTML copy, prompt examples, Sandbox documentation, Open WebUI Models documentation, and monthly usage. Its outline and resources link to Curating knowledge collections as the second session.
Curating knowledge collections
Led by Zach Muhlbauer and Meha Gupta.
Participants clone STEM Adventure Games, choose sources for a topic they want to explore, build their own collection, and revise the cloned system prompt to use those documents. Cloning retains the example's base model and settings; participants spend their time curating sources, revising instructions, and testing.
Workshop Agenda
- Explore and clone · 15 minutes
- Select new sources · 25 minutes
- Create and populate collections · 20 minutes
- Revise system prompts · 10 minutes
- Test and revise · 20 minutes
Lesson Plan
| Minutes | Activity | Evidence to examine |
|---|---|---|
| 0–15 | Introduce RAG, explore STEM Adventure Games, clone and rename it, and try an initial interaction. | Saved copy and an initial response |
| 15–40 | View the example collection. Choose a purpose and audience, then locate three Wikipedia pages for an adapted version. Read relevant passages and keep the URLs. | Participant-selected webpages and intended use for each |
| 40–60 | Create a private collection and a Wikipedia directory. Open the directory and use Add text content to paste substantial excerpts (several paragraphs or complete sections) and source URLs into three separate files. Check saved text, then create instructions.txt at the collection root with directions for using each source. | Three files of excerpts with source links and instructions.txt |
| 60–70 | Attach the new collection to the clone. Edit Purpose and Audience, then replace source and retrieval directions above the numbered game steps with directions to consult instructions.txt. Retain numbered steps, Constraints, and Format, and save. | Prompt instructions that match the new source material |
| 70–90 | Start a blank chat, select the saved copy before sending a request, and choose an adventure. Check citations in that chat. Open the original in another tab to compare the same opening request. Revise one source file or instruction, save the change, then test the saved copy in a new chat. | Source-supported details, gaps, and effects of customization |
The original collection is shown before the customization activity. Newly created collections contain no sources until participants add source text. Do not imply that cloning creates personal copies of shared sources. Participants replace the inherited attachment in their own model and paste Wikipedia articles into their own collection.
Settings are prepared on the example cards before participants clone them. The participant deck does not walk through Function Calling, Advanced Params, or capability switches. Keep custom Skills and Tools unselected; the third workshop introduces those separately.
The STEM opening menu remains immediate. After a selection, the model consults attached documents. Participants put source-specific directions in instructions.txt and replace the source and retrieval directions in their cloned prompt. Remove the inherited one-retrieval-call limit from that replaced block so reading instructions.txt can be followed by retrieving source passages. They can retain, replace, or supplement source ideas from the original game without being assigned its original files.
On September 29, directory creation and saving instructions.txt succeeded through a non-admin account. Add webpage returned “Error processing URL” for three Wikipedia URLs together and for Scientific method alone. This workshop therefore uses copied excerpts and source URLs through Add text content. Live Wikipedia retrieval with tools is deferred to the next workshop. Retrieval through the new instructions.txt workflow remains unverified. Clone-access verification remains pending. Earlier STEM source checking identified an unsupported location, so historical details still need to be checked against quoted passages.
Next steps
- Choose another topic or experiment
- Update copied excerpts and instructions
- Check responses against saved sources
Configuring skills and tools
Thursday, October 15, 2026 · 2:30–4:00 p.m. Eastern · GC New Media Lab 7388.01
Use 13 slides, participant exercises, and complete session notes. Continue participants' existing models and their actual system prompts. Those prompts define adventure setting, goals, and rules. Reusable tools and skills contain no fixed story or genre
Lesson Plan
| Minutes | Time Eastern | Action | Slides |
|---|---|---|---|
| 0–10 | 2:30–2:40 | Continue saved model and explain skill procedures | 1–3 |
| 10–20 | 2:40–2:50 | Add and test skill | 4–5 |
| 20–40 | 2:50–3:10 | Explain, add, and test Adventure Kit | 6–9 |
| 40–50 | 3:10–3:20 | Choose revision or new item and copy content | 10 |
| 50–70 | 3:20–3:40 | Work with CAIL Workspace Support | 11 |
| 70–90 | 3:40–4:00 | Add chosen item, compare results, and reflect across series | 12–13 |
Skill explanation, addition and testing precede tool explanation, addition and testing. After both, participants revise either resource or create their own tool or skill in the same Support activity. Preserve originals and compare matching prompts, models, sources and settings
Conclude with participants' next teaching or research task, changes in their approach across the series, and what they would develop next with prompts, knowledge collections, skills and tools
Preparation records document an authorized Adventure Kit 2.0.0 save on October 14 with ID, ownership and Private access unchanged. An archive fixture was checked in a hosted QA chat. These records do not establish participant access or full acceptance against the standard STEM Adventure Games system prompt. Generic skill comparison and Support roundtrip remain unverified. This website publication installs no live tool and changes no permissions
Source Documentation
Interface instructions draw on the published Sandbox documentation, especially Getting Started, Quick Tour, Models, Knowledge Bases, Tools & Skills, and Roles & Permissions.
The first workshop interface was inspected in Firefox on September 13–17, 2026. Knowledge workshop screenshots and model tests were refreshed on September 28, 2026. Creation, Clone, and access controls were checked in an administrator account. Participant-account checks remain with Zach before delivery. Download Screenshot provenance records source hashes and crop coordinates. Download Showcase provenance distinguishes archival comparison excerpts from current interface instructions. The unrelated fourth screenshot is excluded.
Provider requests are described in the docs as configured for zero retention with training use prohibited. Sandbox history can still be stored and visible to administrators or its shared audience. Retrieved passages enter the model request and may appear in its response. Use materials appropriate for those conditions.
Open WebUI’s Models, Knowledge, and Skills documentation supports the descriptions of custom configurations, retrieval, and skill loading. The Sandbox docs govern local access and sign-in instructions.
STEM Configuration
STEM Adventure Games is the introductory model for Workshops 1 and 2. Its system prompt opens with four experiment options and waits for a choice before retrieving source passages. It presents four numbered choices at each stage and responds to actions typed in ordinary language. The model uses attached sources when discussing historical evidence. It has no attached skills or tools.
STEM Adventure Games — Advanced is reserved for Workshop 3. It uses STEM Adventure and Extend STEM Adventures with native function calling. Its system prompt describes those components and requires a submitted play record before interpreting actions taken inside the game interface.
The original collection contained three Wikipedia imports. List of experiments initially contained a Wikimedia error; complete article text was restored on September 17, 2026, as documented in the Download repair record. Scientific method and Women in science remain alongside it. Additional entries cover Newton’s optical experiments, procedural variants, and software checks. Verify retrieval separately from successful text import.
Additional Knowledge Entries
| Entry | Use |
|---|---|
| Newton: Light and Colour | Check apparatus and distinguish historical claims from game simplifications |
| Newton: Experimental Variants | Support an aperture comparison with a specific source |
| Evaluate Game Procedures | Check commands, prerequisites, replay, and interpretation limits |
| STEM Source Register | Track provenance and document the repaired Wikipedia import |
The Newton entries summarize primary accounts from the Newton Project. They identify their sources and limits; they are not full article imports. Keep Web Search unchanged during comparisons and record whether evidence came from uploaded entries or external pages.
Optional References
Source examples preserve research and historical alternatives. Skill and tool examples preserve the blank template, source interpretation exercise, creator failures, corrections, and test labels. These ordinary pages contain no hidden notes.