Customizing Skills & Tools

A Workshop for the CUNY AI Lab Sandbox

March 30, 2026

Developed by Stefano Morello and Zach Muhlbauer

Workshop Series

Three Weeks, Three Skills

March 16

Composing System Prompts ✓

Defined how the AI thinks, responds, and engages with students

March 23

Curating Knowledge Collections ✓

Grounded the model in course materials so it can reference real documents

March 30 (This Week)

Customizing Skills & Tools

Build specialized skills, tools, and workflows tailored to your courses

Quick Check

Pre-Flight

Verify your setup, then identify the move you’ll build today.

  1. System Prompt — What role and constraints did you set? Have you revised or tested it?
  2. Knowledge Collection — Is it attached? Does the model draw on it when you test?
  3. Name One Pedagogical Move — Something you do as an instructor that follows a repeatable pattern (e.g., narrowing a research topic one question at a time, walking a student through a source, guiding them from description to interpretation).
    Skills enable the model to mimic those moves.
The Basics

What Is a Skill?

A skill is a structured procedure you write in markdown or plain text. It tells the model how to handle a task your system prompt can’t cover on its own.

Think of it as a recipe for a move you’d make as an instructor — establishing stasis, sourcing a document, reading a cinematic frame — written so the model executes it consistently.

Key distinction: The system prompt sets general behavior. A skill decomposes a specific task into steps the model won’t follow unprompted.

What you build
Week 1
System Prompt
behavior & constraints
+
Week 2
Knowledge
syllabus, rubric…
+
New
Skill
step-by-step recipe
student never sees above
Student “Can you look at this paragraph from my draft?”
AI Model
Follows the skill’s procedure to respond
Response “Here’s what I notice in your passage…”
(quotes the excerpt, responds below it)
What the student sees
Skills in Open WebUI

Activating Skills

A skill is worth making when the move has a repeatable structure (trigger, steps, output shape) that you want the model to follow every time.

1. Create the skill

Go to Workspace → Skills and click + New Skill. Write or paste your skill text.

2. Add it to your model

Open your model card in Workspace → Models. Scroll to the Skills section and toggle the skill on. It activates automatically in every chat with that model.

3. Reference it in your system prompt

At the bottom of your system prompt, tell the model which skills to follow — e.g., “Follow the Stasis Theory skill when a student is narrowing a topic.”

Alternative: Paste into System Prompt

You can append skill text directly to your system prompt, but this crowds the prompt and makes it harder to maintain.

Why Bother?

Why Skills Matter

A student types: “I want to write about social media and mental health.”

Without a Skill

“That’s a great topic! You could narrow it by focusing on a specific platform, age group, or type of mental health impact. Consider looking at recent studies on Instagram and anxiety, or TikTok and body image. You’ll also want a clear thesis statement.”

Lists options instead of asking questions. The student still has no method for choosing among them.

With a Stasis Theory Skill

“Your topic: social media and mental health.

Let’s start with the facts. What specific situation are you responding to — something you’ve read, noticed, or experienced? What has happened that makes this a question worth asking right now?”

Opens with one stasis question (conjecture), grounded in the student’s own words. The skill told the model how to narrow, not just to narrow.

The Problem

The Gap Skills Fill

Models can do many things, but not unless you spell it out. Certain moves need a step-by-step recipe to work consistently.

What models won’t do without a skill:

  • Walk through stasis theory one question at a time
  • Quote a passage from an uploaded source before asking a question
  • Ask one focused question instead of listing five
  • Build on what the student just said
  • Scaffold from description to analysis to interpretation
System Prompt Alone
“Help students develop research topics”
Model knows what to do, not how
System Prompt + Skill
“Ask one stasis question, build on the student’s last answer”
Model knows what and how
Tools in Open WebUI

What Is a Tool?

A tool is a function the model can call to do something it can’t do with language alone — search the web, query your uploaded files, run code, or generate an image.

Some tools are built in (Web Search, Code Interpreter). Others are custom: written in Python or imported from the community library via Workspace → Tools. There are plenty of community-built tools available — browse what others have shared.

You can enable tools two ways:

Per Chat

Click the controls icon (sliders, next to +) next to the message bar (1) to open the tools menu (2). Toggle individual tools on for that conversation.

Per Model

Open your model card in Workspace → Models. Scroll to the Tools section to attach custom tools permanently.

Caveat: Smaller models (e.g. Gemma 3 27B) call tools inconsistently. Use Kimi K2.5 or GLM 5 for reliable tool use.

The Difference

Skills vs. Tools

A skill is a document you write in plain text: a procedure the model follows. A tool is code: a function the model calls. Both extend what your custom model can do, but they work differently.

Skill

You write it. Plain text, no code. Tells the model how to handle a specific pedagogical move step by step.

Tool

Pre-built or imported. Code that runs behind the scenes. Gives the model abilities it doesn’t have on its own — like searching the web or running a calculation.

Today’s focus: Skills. Tools are ready to use out of the box. Skills are what you build.

Skill (plain text)
“Walk through stasis one question at a time”
You write it • Shapes how the model responds
vs.
Tool (code)
Search the web, query files, run code, generate images
Pre-built • Gives the model new capabilities
Built-in tools
Web Search
Current information
Knowledge Query
Search uploaded files
Code Execution
Run code in a sandbox
Image Generation
Create from descriptions

Establishing Stasis

Composition — Stasis Theory

Composition & Writing

Before

Starting PointWhen a student is developing a research topic, walk them through four stages — conjecture, definition, quality, and policy — to help them narrow their question. Ask one stasis at a time.
  1. Model walks through all four stages in a single response instead of pausing at each
  2. No procedure for connecting the student’s working topic to each stasis question
  3. Treats stasis as a checklist rather than a deliberative process
  4. No mechanism to let the student reformulate before moving on
  5. Skips the key move: helping the student see which stasis their argument lives in
Composition & Writing

After

StrongSkill: Establishing Stasis for a Research Topic When a student is developing or narrowing a research topic, follow this procedure: Procedure: 1. Ask the student to state their topic in one sentence. Do not evaluate or refine it yet. 2. Conjecture: “What has happened or is happening that makes this worth investigating?” Wait for their answer. Use what they say to sharpen the next question. 3. Definition: Point to a key term in their response. “How are you defining [term]? What kind of problem is this — legal, ethical, empirical, cultural?” Wait. 4. Quality: “What’s at stake, and for whom? What makes this serious enough to argue about in an 8-page paper?” If their scope is too broad, ask them to name one specific population or context. Wait. 5. Policy: “What should be done, and by whom?” Help the student see whether their argument is making a factual claim, a definitional claim, a value judgment, or a policy proposal. 6. Ask: “Which of these four questions does your argument most need to answer?” Guide them toward a thesis grounded in that stasis. Format: Your topic: [student’s stated topic] [One stasis question, tied to a specific phrase the student used] [1–2 sentences explaining why this question matters for their project]

Sourcing a Document

History — The Sourcing Heuristic

History

Before

Starting PointWhen a student asks about a primary source, retrieve it from the knowledge collection and walk them through its rhetorical situation using SOAPS. Ask questions one element at a time rather than summarizing.
  1. Model paraphrases the source instead of quoting from the uploaded document
  2. No procedure for retrieving and presenting specific passages as evidence
  3. Rushes through all SOAPS dimensions in a single response
  4. Student receives a finished reading rather than a structured inquiry
  5. No requirement to ground each analytical move in the source’s own language
History

After

StrongSkill: Sourcing a Primary Document When a student asks about or encounters a primary source from the course, follow this procedure: Procedure: 1. Retrieve the document from the knowledge collection. Quote a key passage — do not paraphrase or summarize. 2. Present the passage in a block quote with its metadata (title, date, author) drawn from the uploaded file. 3. Ask: “Who created this document, and what was their position or stake?” Wait for the student’s answer. 4. After they respond, point to a specific phrase in the quoted passage that supports, complicates, or challenges their answer. 5. Ask: “When and where was this written? What was happening at that moment that shaped what the author could say?” Wait. 6. Ask: “Who was the intended audience? How does knowing that change what the document means?” Wait. 7. After all three sourcing moves, ask: “Given what you now know about the author, the moment, and the audience — what can this source tell us, and what can’t it?” Format: > [quoted passage from uploaded source] — [Author], [Title], [Date] [One sourcing question] [1–2 sentences connecting the question to a specific phrase in the passage]

Reading the Frame

Literature — Cinematic Mise-en-Scène

Literature & Cultural Studies

Before

Starting PointWhen a student shares a film still or visual artifact, guide them from describing formal elements — composition, lighting, framing — toward interpreting how those choices construct meaning in context.
  1. Model describes the image for the student instead of directing their attention
  2. No scaffolding from observation to formal analysis to interpretive claim
  3. Treats all visual elements at once rather than isolating one per turn
  4. No mechanism to keep the student doing the looking and the arguing
  5. Skips the gap between “what’s in the frame” and “what argument it makes”
Literature & Cultural Studies

After

StrongSkill: Reading Cinematic Images When a student uploads a film still, photograph, or visual artifact — or asks about an image from the knowledge collection — follow this procedure: Procedure: 1. If the student uploaded an image, use it directly. If they reference a visual from the course, retrieve it from the knowledge collection and present it. If the image is not in the collection, say so. 2. Ask: “What do you notice first?” Let the student describe before you respond. 3. After their description, direct attention to one formal element they haven’t mentioned — composition, lighting, color, framing, depth of field, or gaze. Ask what it does. 4. Ask how that formal choice shapes the viewer’s experience. Move from what is in the frame to how the image is constructed. 5. Introduce context: ask the student to connect the visual choices to the cultural moment, genre, or argument of the work. If relevant context exists in the knowledge collection, quote it. 6. Guide them toward an interpretive claim: “Based on what you’ve observed, what argument is this image making?” Framework: Description → Analysis → Interpretation • Description: What is literally in the frame? • Analysis: How do formal elements (light, angle, placement) create meaning? • Interpretation: What claim can the student make, grounded in visual evidence?

Writing Your Own Skills

Structure

Anatomy of a Skill

Every skill has three parts.

  1. Trigger — When should this skill activate?
  2. Procedure — What steps does the model follow, in order?
  3. Format — What should the output look like?
Component 1

Trigger

Define when this skill should activate. A clear trigger keeps the model from applying the wrong procedure to the wrong task.

  • What student action starts this workflow?
  • Does it activate when they share a draft? Ask about a source? Upload an image?
  • Should it run automatically, or only when the student asks?
Skill: [Skill Name]
When a student [specific action or input], follow this procedure:

Your turn: What pedagogical move are you trying to teach the model? Name the student action that should trigger it.

Component 2

Procedure

The core of every skill. Numbered steps that tell the model what to do, in what order, and when to wait.

  • What should happen first? What comes next?
  • Where should the model wait for the student before continuing?
  • Should it quote, cite, or reference specific materials?
Procedure:
1. [First step — what does the model do or ask?]
2. [After the student responds, what comes next?]
3. [Continue the sequence — include wait points]
4. [Final step — synthesis, next action, or handoff]

Your turn: Write 3–5 numbered steps. Think about the sequence you follow when you do this yourself as an instructor.

Component 3

Format

Specify what the output should look like. Without a format, the model structures responses however it wants.

  • Should the model quote the student’s text?
  • Should each response end with a question?
  • How long should a response be?
Format:
> [quoted excerpt from student work or source document]

[1-2 sentences: what you observe]
[One question for the student]

Your turn: Write a short format template. What should each response from the model actually look like?

Write Your First Skill

Pick one pedagogical move from your course and turn it into a step-by-step recipe.

Exercise

Choose Your Move

Which is closest to the skill you want to build?

Establishing Stasis

Narrow a research topic one question at a time

Sourcing a Document

Quote, then ask who, when, for whom

Reading the Frame

Describe → analyze → interpret

Something Else

Any repeatable pedagogical move

Draft It

Write Your Skill

Use the four-part structure to write a skill for the move you chose.

1. Trigger

What student action starts this?

2. Procedure

3–5 numbered steps with wait points.

3. Format

What does each response look like?

Skill: [Name]
When a student [trigger action], follow this procedure:

Procedure:
1. [First step]
2. [Second step — include wait points]
3. [Third step]

Format:
> [quoted text from student or source]

[Observation in 1-2 sentences]
[One question]
Integrate It

Add Your Skill

Connect the skill you wrote to the model you built in the previous workshops. Follow the same steps from slide 5.

  1. Create the skill — Go to Workspace → Skills, click + New Skill, and paste in what you wrote.
  2. Add it to your model — Open your model card in Workspace → Models. Scroll to the Skills section and toggle it on.

Test It

Open a new chat. Trigger the skill with realistic student input. Does it follow the procedure?

Iterate: If the model skips steps, add wait points. If it dumps everything, add “one per turn.”

System Prompt
Role, constraints, tone
Knowledge Collection
Syllabus, rubric, readings, sources
Your Skill
Step-by-step recipe for a specific move
Your Complete AI Tool
Directive + sources + workflows
Workshop Complete

What You’ve Built

March 16

System Prompts ✓

Role, constraints, tone

March 23

Knowledge Collections ✓

Grounded in your materials

March 30 (Today)

Skills & Tools ✓

Step-by-step workflows

You now have a custom AI tool with three modular layers that can be tested as a configuration and refined iteratively.
That way, you can see what each component contributes and where to focus before you introduce it in the classroom.

ailab.gc.cuny.edu