QR code to slides at aud613.niyaziarslan.com

            follow along
        

Human in the Loop

Faculty-Directed AI
for Canvas Course Development

Niyazi Arslan, Ph.D.

Department of Speech Pathology and Audiology

Where teaching lives now

Half of US higher-ed enrollment runs on Canvas.

1996

WebCT

1997

Blackboard

1999

D2L

2002

Moodle

2011

Canvas

2020

Pandemic

2023+

AI in LMS

50%

↑ 44 → 47 → 50

20%

SUNY + CUNY moves · 2024

12%

↓ 21 → 18 → 12 · Anthology Ch. 11 '25

Holding steady · smaller schools

USAonline Sakai → Canvas, spring 2020

                    Phil Hill / On EdTech YE-2024 · Edutechnica spring 2025 · Inside Higher Ed (Anthology Ch. 11)
                

Most of us teach through Canvas now. But the LMS as a category is thirty years old, and the picture has shifted several times. Some of you in the room have probably used different platforms across your career. Let me back up. WebCT was the first online platform built for this purpose in nineteen ninety-six. Since then, different platforms emerged, right? Blackboard, D2L Brightspace, Moodle, Canvas. Finally, twenty-twenty, the pandemic, which made the LMS essential for every faculty member in every discipline, not just for online teaching. And from twenty twenty-three onward, every major LMS started working on integrating AI at the platform layer. That is the overall history. And here is where we are today. As you can see in the chart, Canvas is now larger than the other three combined. Now let me show you what filling it takes, before any student logs in.

What the platform has to hold

Building a full Canvas course.

Home

Announcements

Syllabus

Modules

Pages

Files

Assignments

Quizzes

Rubrics

Discussions

Grades

Panopto Video

Empty course shell. Nothing here yet.

Syllabus · course design

Map alignment

Goal 1 → Quiz 1 → PS 1

Sequence modules

M01 → M02 → M03 → M04

Apply UDL

Engagement Representation Action & expression

Pages · Files · Captioned media

Lecture 01

slides · captions

Lecture 02

slides · captions

Reading 01

PDF

Lab demo

video 14m

Reading 02

PDF

Audio set

.wav x 8

Assignments · Quizzes · Rubrics

Problem set 1 · with rubric 50 pts

Quiz 1 · 20 items, banked 20 pts

Lab report 1 · peer review 100 pts

Problem set 2 · with rubric 50 pts

Modules · everything in order

M01Sound and waves

Lecture 01 · Reading 01

Problem set 1

Quiz 1

M02Hearing and threshold

Lecture 02 · Reading 02

Audio set · Lab demo

Lab report 1 · Problem set 2

Week 6 · live operations

Announcement posted · "Week 6 reminders" 2h ago

22 submissions awaiting grade PS-2

Discussion · "Threshold interpretation" 14 posts

01 Design

Map alignmentSequence modulesApply UDL

02 Build content

Record lecturesBuild slidesCurate readingsCaption media

03 Build assessments

Write itemsBuild rubricsQuestion banksPeer review

04 Publish to Canvas

Build modulesCreate pagesAdd alt textAudit accessibility

05 Run the term

Weekly announcementsGrade submissionsFeedbackFacilitate discussion

Let's look at the different kinds of work that go into building a full Canvas course. Stage one is design. This is the pedagogical work that happens before any clicking, mapping alignment, sequencing modules, applying universal design. Stage two is building content. Recording the lectures, building the slides, curating the readings, captioning the media. For a fifteen-week graduate course, this turns into dozens of pages, dozens of files, and hours of captioning that has to be accurate. Stage three is building assessments. Writing the items, building the rubrics, setting up question banks. Stage four is publishing to Canvas. Building the modules, creating the pages, handling accessibility. Finally, stage five is running the term. Once the course is live, announcements go out every week, submissions need grading, discussions need facilitating. This is the part of the work that does not end until the semester does. However, students see only stage five. Administrators see only stage five. Faculty do all of this, every semester, for every course they teach. Let me show you what this work actually costs.

The work behind a good Canvas course

Hidden behind every stage.

01

Time

70hrs

Median to build a first online course. Then 12.69 hrs/wk to run it. Scales per student: ~30 min/week, each.

                        Freeman 2015 (OJDLA)

                        Mandernach & Holbeck 2016

02

Expertise gap

28%

of faculty fully prepared for online course design. Flat since 2020.

                        CHLOE 10 (QM & EDUCAUSE)

                        n=257 COLOs · Jan-Feb 2025

03

Design

300hrs

to reach beginner HTML/CSS, per freeCodeCamp. That's the floor for escaping Canvas's editor, which strips JavaScript, custom CSS, SVG, and animation on save. 700+ institutions pay Cidi Labs to bridge the gap.

                        canvas_sanitize.rb

                        freeCodeCamp · Cidi Labs

04

Accessibility

Apr 26

2027

federal deadline

DOJ Title II. WCAG 2.1 AA on every Canvas page, every PDF, every video. 95.9% of pages on the web fail today.

                        DOJ Title II · HHS OCR

                        WebAIM Million 2026

05

Maintenance

38%

of webpages from 10 years ago no longer load. The course is the same. The dates, rules, and links are not.

                        Pew Research 2024

                        “When Online Content Disappears”

These are hidden behind every one of the stages I just walked through. One. Time. The median is seventy hours just to build a first online course. Then about twelve and a half hours per week to run it. And here is the part that gets you. Roughly half of that weekly work scales linearly with class size, about thirty minutes per student, per week. A section of forty students is twenty hours a week. Multiply by your teaching load. Two. The expertise gap. The research shows that only twenty-eight percent of faculty feel fully prepared for online course design. That number has been flat since the pandemic. Three. Design. Canvas's editor is not a free environment. By design, it strips JavaScript, custom CSS, and animation when you save. So if you want anything interactive on a page, a slider, a tabbed reading, an embedded simulation, you have two real options. Learn HTML and CSS, which freeCodeCamp puts at three hundred hours just to reach the beginner floor. Or pay a vendor. Over seven hundred institutions pay Cidi Labs to bridge that gap. Four. Accessibility. April twenty-six, twenty twenty-seven. The Department of Justice's Title Two final rule, which requires every Canvas page, every PDF, every video at our institution to meet WCAG two point one Level AA. One year from now. And to put the work into context, the WebAIM Million project measured the open web this year. Almost ninety-six percent of pages fail WCAG today. Five. Maintenance. Most of us do not teach a new course every year. We teach the same course again and again. And every time we do, the world around it has shifted. Pew Research measured this in twenty twenty-four. Thirty-eight percent of webpages from ten years ago no longer load. Now look at your Canvas course. The links you embedded five years ago. The video you assigned three years ago. The reading you scanned in twenty nineteen. Some fraction of those are already dead. And it is not just links. The AI policy in the syllabus did not exist in twenty twenty-two. The course is the same. The world around it is not. These do not sit separately. They compound. Time, the expertise gap, the design constraints, the accessibility deadline, the maintenance tax. Every course, every August. That is what every faculty member in this room is dealing with. So the question is the obvious one.

How can AI help with each one?

                01  Time
                02  Expertise
                03  Design
                04  Accessibility
                05  Maintenance
            

Where AI fits.

01

Time

70 hrs to build, 12.69 hrs/wk to run, ~30 min per student per week.

→

AI helps faculty draft the build (modules, pages, quizzes, rubrics from your syllabus) and the weekly work (announcements, feedback, grading rubrics).

02

Expertise gap

28% of faculty fully prepared. Flat since 2020.

→

AI accelerates the Canvas learning curve in real time. The 28% narrows with use.

03

Design

Canvas strips JS, custom CSS, SVG, animation. 300 hrs of HTML/CSS to escape.

→

AI helps faculty build modern, interactive pages. Coding becomes minutes, not hours. No third-party tools needed.

04

Accessibility

WCAG 2.1 AA by Apr 26, 2027. 95.9% of pages fail today.

→

AI helps faculty fix accessibility gaps page by page. Alt text, contrast, headings, captions: each proposed, each approved. The 95.9% drops one page at a time.

05

Maintenance

38% of 10-year-old webpages no longer load. Dates, rules, links all shift.

→

AI helps faculty run the pre-semester audit. Dead links flagged, dates re-anchored, policy text drafted, each one approved before it lands.

Here is where AI fits. One row per burden, one answer for each, and one rule that runs through all of them. One. Time. AI drafts the build for you. Modules, pages, quizzes, rubrics, from your syllabus. And it drafts the weekly work, announcements, feedback, grading rubrics. You edit and publish. Two. The expertise gap. Twenty-eight percent of faculty fully prepared, and five years of workshops have not moved that number. AI accelerates the Canvas learning curve in real time. The twenty-eight percent narrows with use, not with another training. Three. Design. The Canvas editor strips most of what makes a page interactive. AI helps faculty build interactive pages directly. Coding becomes minutes, not hours. No third-party tools needed. Four. Accessibility. The April twenty twenty-seven deadline is real. AI helps faculty fix gaps page by page. Alt text, color contrast, heading hierarchy, captions. Each fix proposed, each waits for your yes. The ninety-six percent of failing pages drops one page at a time. Five. Maintenance. The course is the same. The world around it is not. AI helps faculty run the pre-semester audit. Dead links flagged, dates re-anchored, policy text drafted. Each one waiting for your yes. And here is the rule that runs through all five. AI proposes, faculty says yes or no. Nothing lands without your yes. That is the whole architecture. And yet most of us already use AI for some of this work. So what is different about what I am proposing? Let me show you the gap.

The chat-and-copy pattern

What most faculty do today.

01 · YOU

Faculty

the human in the loop

Does every transfer by hand.

types prompt

reads output

02 · AI (web)

Web AI

ChatGPT · Gemini · Claude.ai

Chat in a browser tab. Never sees your course.

copy

paste · fix format

03 · VENUE

Canvas

USAonline

Real institutional courses.

Every transfer is a tab switch. Can we do this more systematically?

Many of you are already doing some version of this. You open ChatGPT, or Claude.ai, or Gemini in a browser tab. You type a prompt for a Canvas page, a quiz question, an announcement, a feedback comment. The AI gives you something. You select it. You copy it. You switch to your Canvas tab. You paste it into the editor. You fix what broke in the formatting. You switch back to the AI for the next thing. It works. People have built whole courses this way. But notice what is between the AI card and the Canvas card on this slide. Just two arrows labeled copy and paste, and your hand doing the work. There is no bridge between the AI and Canvas. The AI never reads your actual course. It does not know what is in your gradebook, what your modules look like, what your students have submitted. It is guessing from your prompt. You fill in the gaps from memory and from copy-paste. Every transfer is something you do by hand. The format breaks on paste and you fix it every time. And if you want consistency across multiple sections, or across the same task done in different weeks, there is no architecture to support that. You either remember to use the same prompt every time, or you do not. Every transfer is a tab switch. The question is, can we do this more systematically?

How it works

AI inside Canvas, systematically.

01 · YOU

Faculty

the human in the loop

Directs every action.

directs

says yes

02 · AI

Claude

by Anthropic

Claude Code the runtime where this all happens
the systematic layer
Skills Markdown files Claude reads for each task.
Hooks Validators that fire after every Claude write.
GATE protocol Pauses Claude until you say yes.

calls

returns

BRIDGE

canvas-mcp

open-source MCP server

vishalsachdev/canvas-mcp

via MCP

via Canvas REST API

executes

responds

03 · VENUE

Canvas

USAonline

Real institutional courses.

                Components:
                1. Claude Code
                ·
                2. The bridge to Canvas (canvas-mcp)
                ·
                3. Skills · Hooks · GATE
            

Anthropic Claude Code · vishalsachdev/canvas-mcp · open standard (Model Context Protocol)

Yes, we can. Three connections. One rule. On the left, you. You direct every action. Nothing happens without your yes. In the middle, Claude. Anthropic's model, running inside Claude Code. On top of Claude are the three things I built myself: skills, hooks, and the GATE protocol. Between Claude and Canvas is canvas-mcp. Open-source, written by Vishal Sachdev. It translates between two languages: MCP on the Claude side, the Canvas REST API on the Canvas side. This is the bridge that was missing from the previous slide. On the far right, Canvas. USAonline. Real institutional courses, not a sandbox. With the bridge in place, the AI reads your actual modules, your actual gradebook, your actual roster. When it writes a page, the page goes into your course directly. I direct. Claude proposes. I say yes. The change lands. Nothing touches the live course without my yes. Three components, in order. Starting with Claude Code, the runtime.

Component 1 of 3 · the runtime

v2.1.138 · May 2026

Claude Code.

Anthropic's agent. Runs in your terminal or right inside VS Code. Reads your files, calls tools, waits for you.

Browser chat · web application

claude.ai

Help me redesign my Module 6 page in Canvas.

Sure. Paste the page text here and I will suggest edits.

Can't you just open it in Canvas?

No system access from this chat.

Single conversation. Text in, text out. No system access.

Claude Code · desktop application

claude code · aud613

> Help me redesign my Module 6 page in Canvas.
✓ Calling canvas-mcp / get-page
✓ Checking WCAG 2.1: alt text, contrast, headings
I see 3 accessibility issues and have a redesign proposal.
continue / revise / skip ?
> C
✓ Calling canvas-mcp / update-page
✓ Page updated. Refresh Canvas to confirm.

Local files, external tools. Same agent in terminal or VS Code.

Component one. The runtime. The piece where the AI actually sits and works. This is what most of you have used. A browser tab, ChatGPT or Gemini or Claude.ai. You ask, it answers in text. But ask it to actually open your Module 6 page in Canvas, or read your gradebook, or push an announcement. It cannot. No system access. So you copy. You paste. You fix what broke in the formatting. Now this. Same model, different surface. Claude Code is an app you install on your own laptop. I ask the same question. It reads the live page. It checks accessibility on its own. It tells me what it found and pauses. Three options: continue, revise, skip. I type C. The page updates in Canvas. I open Canvas to verify. One clarifier. This is not a local language model. The Claude model itself still lives on Anthropic's servers. What runs on your laptop is the agent app, the one that reads your files and reaches Canvas. You still need internet. The model is in the cloud. The workflow is on your desk. That is component one. Next, the protocol it uses to reach Canvas.

Component 2 of 3 · the bridge to Canvas

MCP. The Model Context Protocol.

An open standard for how AI assistants connect to outside tools and data.

What it is

AI host

Claude Code

also: ChatGPT, Gemini, Copilot

MCP

canvas-mcp → Canvas (USAonline)

github-mcp → GitHub

filesystem-mcp → local files

slack-mcp → Slack

+ 10,000 more · one protocol, many tools

One protocol that connects any AI on one side to any tool on the other.

Component two. The bridge to Canvas. Two slides for two pieces. The open standard first. Then the actual bridge. What is a protocol? Think of a wall outlet. Same prongs on every wall, no matter what you plug in. MCP, the Model Context Protocol, is that outlet for AI tools. Look at the diagram. On the left, the AI. Could be Claude Code, ChatGPT, Gemini, Copilot. On the right, the tools the AI reaches. Your files, GitHub, Slack, Canvas. MCP sits between them, same shape on both sides. Why this matters. Wire each tool up once and any MCP-aware AI can use it. Not twelve times for twelve different AIs. One protocol, ten thousand tools. You will not need to know how MCP works to teach with this workflow. But knowing it exists is what lets the next slide make sense. The highlighted row in the diagram is canvas-mcp. That is next.

Component 2 of 3 · the bridge to Canvas

                    
                    vishalsachdev/canvas-mcp · v1.3.0 · 130+★ · MIT
                

canvas-mcp + the Canvas API.

Canvas exposes a REST API. canvas-mcp turns it into MCP tools Claude can call.

How it connects

Layer 1 · the LMS

Canvas (USAonline)

REST API

Layer 2 · the bridge

canvas-mcp

Python · runs locally · 88 tools, 8 agent skills

MCP

Layer 3 · the AI

Claude Code

88 tools across categories like:

pages modules assignments quizzes discussions files users announcements rubrics analytics

What you need from Canvas

API TOKEN a private key from your Canvas account

                            7~Xa9bC4d2FgH8jK•••••••••••••••••••••••••••••••••••••••••9pQ
                        

FROM CANVAS Account → Settings → + New Access Token

TREAT LIKE A PASSWORD

Never share. Never paste in chat. Anyone holding this can act as you in Canvas.

FERPA

Student names auto-anonymized before leaving your laptop.

Built by Vishal Sachdev, Clinical Associate Professor & MakerLab Director, Gies College of Business, UIUC. github.com/vishalsachdev/canvas-mcp

A note

canvas-mcp is not the only way to reach Canvas. The real interface is the Canvas REST API, which is public.

Now the bridge itself. The piece that translates between Claude and Canvas. Architecture on the left. Three layers. Canvas at top. Claude Code on your laptop at bottom. canvas-mcp in the middle. The connectors show the two languages it speaks: REST API on Canvas's side, MCP on Claude's side. Translating between them is its whole job. Eighty-eight tools live inside it, grouped into the categories you see as pills. Pages, modules, assignments, quizzes, discussions. Verbs Claude Code can run on your behalf. On the right, what you actually need to make this work. One thing. A Canvas API token. Think of it as a private key from your own Canvas account. Account, Settings, New Access Token. Two clicks, no IT ticket. The token inherits your faculty permissions, nothing more. Take the warning seriously. Anyone holding this token can act as you in Canvas. Never share it. Never paste it into a chat window. It belongs in a config file on your laptop and nowhere else. Student privacy is built in. canvas-mcp anonymizes student names before any data leaves your machine. I will come back to that on the FERPA slide. At the bottom of the right column you see who built this. Vishal Sachdev, faculty at UIUC. Open source on GitHub. Faculty, not a vendor. One honest note. canvas-mcp is not the only way to reach Canvas. The Canvas REST API is public. A Python script, a curl command, the canvasapi library, any MCP server would work. I use canvas-mcp because someone else already did the integration work. The architecture does not care which wrapper you pick. Next, your customization layer. Skills, hooks, and gates.

Component 3 of 3 · your customization layer

                    .claude/ · skills + hooks + gates · all markdown, no code
                

Skills, hooks, gates.

Three mechanisms you write yourself. Your standards, your checks, your control.

01 · SKILLS A syllabus written for the AI

claude code · aud647

> Build Module 6 page from the brief

✓ Loading skill: canvas-content-skill
✓ Reading SKILL.md
✓ Loading 3 reference files
  voice-guide.md
  accessibility-rules.md
  course-constants.md

Phase 1 of 4, reading the brief...

.claude/skills/canvas-content-skill/SKILL.md

---
name: canvas-content-skill
description: Build a Canvas page
  from a lecture brief. Apply voice
  and accessibility rules.
---

# Build a Canvas page

## Phase 1: Read the brief
## Phase 2: Apply voice
## Phase 3: Build the page
**GATE:** present, wait for C/R/S/N
## Phase 4: Validate and save

A skill is markdown. You write it once. Claude follows it every time.

02 · HOOKS Auto-validators that fire on every write

claude code · aud647

✓ Phase 3 complete. Wrote pages/module-6.html

[hook: PostToolUse → validate-canvas-html.sh]

× HOOK BLOCKED · 2 issues:
  • image at line 47 missing alt attribute
  • h2 → h4 heading skip at line 52

Reverting write. Fix and retry.

.claude/settings.json

{
  "hooks": {
    "PostToolUse": [{
      "matcher": "Write|Edit",
      "hooks": [{
        "type": "command",
        "command": ".claude/hooks/
          validate-canvas-html.sh"
      }]
    }]
  }
}

A hook is a script you write. Fires on every Claude write. If the file fails your check, Claude can't move on.

03 · GATES Pause for your yes at every phase

claude code · aud647

✓ Phase 3 complete. Draft saved.
✓ Validators passed.

Phase 3 → Phase 4 boundary

GATE: How would you like to proceed?

  (C) Continue to Phase 4
  (R) Revise (give feedback, re-gate)
  (S) Skip this phase
  (N) Next phase only

> _

.claude/skills/canvas-content-skill/SKILL.md

## Phase 3: Build the page

Generate the page using the
template. Apply voice and
accessibility rules. Run the
validator before presenting.

**GATE:** Present the draft to
the user. Wait for one of:
  (C) Continue to Phase 4
  (R) Revise (apply feedback, re-gate)
  (S) Skip this phase
  (N) Next phase only

GATE is something you build using skills. Your stop sign at every phase boundary.

Web AI + copy-paste

Standards live in your head
Format breaks on every paste
You're in the loop only at the prompt
Re-prompt every session

This workflow

Standards live in markdown on disk
Hooks validate every write
You're in the loop at every phase
One skill, every course

You write the standards once. They run on every page, every course, every term.

Component three. Your customization layer. Where you write your own rules, checks, and decision points. Three mechanisms, one at a time. First, skills. Think of a skill the way you think of a checklist you hand to a teaching assistant. Step by step, here is how a task gets done in our course. A skill is the same thing, written for Claude instead of for a TA. It is a markdown file. On the right, what one looks like. YAML at the top tells Claude when to use it. Below that, ordered phases in plain prose. On the left, what happens when I ask Claude to build a Canvas page. The skill loads, the references load, phase one begins. Other skills you might write: build a quiz from objectives, draft a weekly announcement in your voice, generate a rubric from a template you already use. Second, hooks. Think of a hook like a spell-checker that runs every time you save. Define the rule once, it runs forever. A hook is a small script that fires every time Claude writes or edits a file. On the right is the configuration in settings dot J S O N. The matcher says which tools to fire on. On the left, what happens when Claude writes a page that fails your accessibility check. The hook blocks the write. Claude cannot move on until the file passes. Enforcement that runs without you lifting a finger. Third, gates. Think of a gate like a stop sign at every intersection. The AI has to pause and wait for you to wave it through. GATE is a convention you write into every phase of every skill. On the right, the SKILL dot M D excerpt: each phase ends with a GATE directive that tells Claude to present the work and wait. On the left, what you see in your terminal. Four letters. C continue, R revise, S skip, N next phase only. You type one. The decision stays yours, every phase. Compare back to the chat-and-copy pattern. With web AI, your standards live in your head. Format breaks on every paste. You are in the loop only at the prompt. You re-prompt every session. With skills, hooks, and gates, your standards live in markdown on disk. Hooks validate every write. You stay in the loop at every phase. One skill works on every course you teach. You write the standards once. They run on every page, every course, every term. Now let me show you what this looks like in practice.

Putting it all together

The full loop.

What this looks like in Claude Code.

Claude Code · aud647 · VS Code

canvas-content-skill loaded

                            > Build the Module 6 page from the brief.
                        
✓ Loaded skill canvas-content-skill (4 phases, 3 reference files)
✓ Applied voice-guide.md and accessibility-rules.md
Phase 1: read brief  ·  done
Phase 2: apply voice  ·  done
Phase 3: build page draft  ·  ready for review
GATE · Phase 3 boundary
Draft ready. How would you like to proceed?
  (C) Continue to Phase 4 (validate)
  (R) Revise (give feedback, re-gate)
  (S) Skip this phase
  (N) Next phase only

                            > C   # continue
                        
✓ Writing pages/module-6.html
[hook fired: PostToolUse → validate-canvas-html.sh]
✓ Alt text: 8 of 8 images covered
✓ Heading hierarchy: H2 → H3 valid
✓ Link integrity: 12 of 12 resolve
✓ Hook passed
✓ Calling canvas-mcp → update_page
✓ Pushed to Canvas REST API
✓ Module 6 page updated in your live course.
Open USAonline to verify.

Let me show you what this looks like end to end. The screen is what I see in my VS Code window when I run this workflow. One task, all three components, in sequence. Step one. I type my request in the Claude Code panel. Build the Module 6 page from the brief. Step two. Claude loads my canvas-content-skill, reads the four phases, and applies my voice and accessibility reference files. Phases one and two run automatically. Phase three produces a draft and stops. Step three. The pause. The GATE I wrote into phase three fires. Claude presents the four options. C, R, S, N. Waiting for me. Step four. I type C. The deciding stays mine, even when the work is good. Step five. Claude writes the file. The PostToolUse hook fires. Alt text on every image, heading hierarchy, link integrity. All pass. Step six. canvas-mcp translates the file to MCP, calls the Canvas REST API, and pushes the page to my live USAonline course. Done. I open Canvas to verify. One task. All three components touched. The labor of writing, validating, and pushing has moved to the system. The deciding, at every gate, stays yours. Now let me show you what this same loop can do across a full course.

The payoff · what this lets you do

What you can do on Canvas with this.

Different tools.

BUILD module · announcement · page when there is nothing yet

claude code · aud613

> Build Module 6 from the AUD613 syllabus

✓ Loading canvas-content-skill
✓ Phase 1-3 complete · draft ready

GATE: Continue?
> C

✓ Created 5 pages, 5 assignments
✓ Created 1 announcement
✓ Pushed to USAonline

AUD-613-101 · Psychoacoustics

Home

Announcements

Modules

Pages

Assignments

Grades

Modules

Module 6: Sound Localization

✓ Just created · 5 pages, 5 assignments

• Lecture 6.1: Interaural cues

• Lecture 6.2: Cone of confusion

• Lecture 6.3: HRTFs

° Problem Set 6: due Mar 20

° Quiz 6: due Mar 25

MAINTAIN due dates · links · structure when the calendar moves underneath

claude code · aud647

> Roll all due dates: Spring 2026 to Fall 2026

✓ Scanning 47 assignments, 12 module unlocks
✓ Preserving day-of-week alignment

GATE: 59 date changes ready. Apply?
> C

✓ Updated 47 due dates
✓ Updated 12 module unlocks
✓ Pushed to USAonline

AUD-647-101 · Audiologic Rehab

Home

Announcements

Modules

Pages

Assignments

Grades

Assignments ✓ 47 dates rolled

Audiogram interpretation Aug 28

Hearing aid fitting brief Sep 11

Counseling roleplay Sep 25

Pediatric case study Oct 9

+ 43 more

AUDIT accessibility · triage · coverage when the question is what is wrong

claude code · audit

> Audit all pages for WCAG 2.1 AA

✓ Scanning 133 pages across both courses
✓ Checks: alt text, contrast, headings, links

GATE: Show report?
> C

! 12 critical · missing alt text
! 23 medium · heading order
! 8 low · link text quality

AUD-613 + AUD-647 · Accessibility report

Home

Modules

Pages

Files

Quizzes

Grades

Pages with issues 43 of 133 flagged

Module 3 / Lecture 3.2 CRITICAL · 3 issues

Module 5 / Audiogram intro CRITICAL · 2 issues

Module 7 / Frequency response MEDIUM · 4 issues

Module 9 / Speech audiometry MEDIUM · 2 issues

Module 11 / Pediatric assessment LOW · 1 issue

+ 38 more

A sample of canvas-mcp's 88 tools github.com/vishalsachdev/canvas-mcp

Read & browse

list_courses list_assignments get_course_details list_submissions

Build

create_module create_page create_assignment create_announcement

Communicate & grade

send_conversation bulk_grade_submissions grade_submission list_discussion_topics

Bulk & files

upload_course_file execute_typescript enroll_user create_rubric

+ 72 more. Any tool, in any order, composed by skills you write.

Three modes you will use this in. Build, maintain, audit. Same loop, different verbs. First, build. When there is nothing yet. On the left, my Claude Code panel. I ask Claude to build Module 6 from the syllabus. The canvas-content-skill loads. Phases run. I say C at the GATE. canvas-mcp pushes five pages, five assignments, and one announcement to USAonline. On the right, what that looks like in Canvas. Module 6 expanded, pages nested inside, all in a single loop. The same skill builds a weekly announcement in your voice, or any single page from raw notes. Second, maintain. When the course already exists and the calendar moves underneath it. I ask Claude to roll every due date from Spring twenty twenty-six to Fall twenty twenty-six. canvas-mcp scans forty-seven assignments and twelve module unlocks, calculates Fall offsets, presents the plan. I say C. The changes land. The same skill can find and replace broken external links, or standardize page titles across modules. Third, audit. When the question is not what to add but what is wrong. I ask Claude to audit every page in both courses against WCAG two point one AA. One hundred thirty-three pages scanned in one pass. Twelve critical, twenty-three medium, eight low. On the right, the pages list with severity flags. I work through them one at a time, each fix gated. The same skill can generate a triage report, or cross-check coverage against the syllabus. At the bottom of the slide, a sample of canvas-mcp's actual tool names. Eighty-eight in total, grouped four ways. Read, build, communicate and grade, bulk and files. Skills compose any of them in any order you write down. Now let me show you the most unexpected place this loop can run from.

When you're not at your desk

                    Claude mobile app · Remote Control · canvas-mcp
                

From your phone, by voice.

01 · Dictate

“Post an announcement to AUD647: Friday's lab moves to Room 312.”

iOS keyboard dictation (system mic)

02 · Relay

Phone routes to your laptop session, in seconds.

claude.ai/code Remote Control · laptop awake

03 · Draft + Gate

Skill drafts on the laptop. The C/R/S/N gate mirrors to your phone.

build-announcements · canvas-mcp loaded

04 · Publish

Live in Canvas. Email + push to all enrolled.

canvas-mcp · pre-approved API permission

The most unexpected place this loop can run from. Your phone. With your laptop somewhere else. Picture this. I am between clinic visits. My phone is in my pocket. My laptop is back at the office, still on. I remember that Friday's lab needs to move. The room is being used for a conference. No time to get back to the office. Step one. I open the Claude app on my phone, tap into the Code tab, pick the AUD647 project. Claude Code came to iOS and Android in October twenty twenty-five. I tap the microphone on the iOS keyboard and dictate one sentence. Post an announcement to AUD six four seven, Friday's lab moves to Room three twelve. Step two. That request routes through Anthropic's Remote Control feature, shipped on February twenty-fifth, twenty twenty-six. Remote Control opens a window from my phone into the Claude Code session running on my laptop. The laptop stays awake. canvas-mcp and the build-announcements skill are already loaded. Step three. Claude drafts the announcement. Title, body, course context, sign-off. The same GATE you saw on earlier slides shows up. C, R, S, N. The GATE mirrors to my phone, so I see exactly what is about to publish. Step four. I tap C on my phone. canvas-mcp publishes. Twenty-two students get an email. Twenty-two get a push notification. Done. Two things before we move on. On FERPA. The example I gave is logistics, a room change. Directory information at most. If the announcement needed to name a student, I would not dictate it in a public place. Thirty-four C F R ninety-nine point three counts oral disclosure as a FERPA event. The GATE is non-negotiable. Even on the phone, nothing publishes until I type C. Canvas has no announcement recall. Once it flushes to email, you cannot pull it back. The GATE is what makes this workflow safe to do at all. Same Claude Code session. Same skill files. Same C R S N. The room changes. The GATE does not.

Walkthrough · one prompt, one announcement

                    AUD647 · build-announcements skill · canvas-mcp
                

One real prompt. One real Canvas page.

The terminal on the left replays how this announcement got built. The screenshot on the right is what landed in Canvas.

claude · AUD647 / spring 2026

$ asciinema cast loading...

Serve the deck via local HTTP for the cast to load:
cd presentations && python -m http.server

asciinema replay · build-announcements skill REPLAY

Canvas · AUD647 announcement

Study Guide Is Ready! announcement on Canvas

USAonline · Study Guide Is Ready! · published SHIPPED

Let me walk you through one of these end to end. This is the AUD647 midterm study guide announcement that went out earlier this semester. On the left is the prompt I gave Claude, replayed in the terminal. Watch as it types. Midterm study guide is up. Exam is Monday March nine, in class. Covers Module 1 Lecture 1 through Module 4 Lecture 1. Needs Respondus LockDown Browser. Forty-five characters of intent. That is all I had to write. Watch what happens next. The build-announcements skill loads. Claude drafts the title, Study Guide Is Ready exclamation point, in my voice not generic LMS voice. Then the body, in the structure I always use for these. A friendly opening, the exam scope, the date and format, the technical requirement. Then the gate. C, R, S, N. I read the draft, I type C. canvas-mcp posts the announcement to AUD647. Twelve students get an email. The screenshot on the right is the actual announcement that landed in Canvas. Same title. Same friendly opening. Same exam details, same Respondus note. The skill produced what the skill was designed to produce. Two things I want you to take from this. One, the gate fires every single time. The only thing on screen that I personally typed was the original sentence and the letter C. Two, the screenshot on the right is not a mockup. That announcement is in the AUD647 Canvas course right now. Twelve students read it. Now, what does this add up to.

what shipped this term through the setup

                    AUD613 (57 pages) · AUD647 (48 pages) · 26 announcements
                

AUD613 + AUD647 · Spring 2026.

click to enlarge AUD613 · Home

Psychoacoustics · Course home module roadmap

click to enlarge AUD613 · Announcement

Welcome to Psychoacoustics term opener

click to enlarge AUD613 · Lecture

Physics of Sound · M01 L01 lecture overview

click to enlarge AUD647 · Home

Adult Rehabilitation & Aging course home

click to enlarge AUD647 · Announcement

Midterm study guide is ready prep nudge

click to enlarge AUD613 · Assessment

Module Test 1 · Logistics exam page

The workflow, piece by piece. Here is what it shipped across a whole term. Two courses, Spring twenty twenty-six. AUD six one three, Psychoacoustics. Fifty-seven Canvas pages. AUD six four seven, Adult Rehabilitation. Forty-eight pages. Twenty-six announcements between them. Every one drafted by Claude through one of my skill files, walked through the same GATE, pushed through canvas-mcp. These six are not screenshots. They are the actual Canvas pages, embedded live. I can click any of them to enlarge. Top row, AUD six one three. Course home with the wave illustration and module roadmap. The welcome announcement that went out on day one. A lecture overview for the physics of sound. Bottom row, AUD six four seven. Home page for Adult Rehabilitation and Aging. The midterm study guide announcement. The module test logistics page with the Respondus banner. Six tiles, the whole shape of what runs through this setup. Course shells. Weekly announcements. Lecture pages. Assessment logistics. Different audiences, different verbs. Same skill files, same GATE, same canvas-mcp. I am not putting up time savings numbers. I do not have clean enough records to back them up, and the honest answer is it depends on the page. What I can put up is this. Every one of these pages is in Canvas right now. Every one got my yes, line by line, before it shipped. Which leads to the next question. If AI is drafting course pages, and if any part of this ever touches student work, what happens with FERPA.

FERPA · how it's protected in this workflow

                    canvas-mcp · anonymize_response_data()
                

Two zones. Two answers.

When AI helps with course pages vs when AI helps with student work, the protection looks different.

Zone 1 · Authoring

No student data enters.

Pages, modules, syllabi, announcements. The AI works on artifacts, not on people.

pages modules syllabus announcements files settings

Nothing the AI sees identifies a student. FERPA is not in scope for this zone.

✓ every endpoint outside /users, /submissions, /discussion, /enrollments, /groups, /analytics

Zone 2 · Student data

canvas-mcp anonymizes at the source.

Grading, summarizing, analytics. The AI sees the work, not the name. Identifiers are hashed before the data leaves your machine.

Canvas response

id: 12345

name: Maria Rivera

email: mrivera@usouthal.edu

login_id: mrivera42

sis_user_id: J00123456

canvas
-mcp

What Claude sees

id: 12345

name: Student_5994471a

email: student_5994471a@example.edu

login_id: student_5994471a

sis_user_id: null

Also redacted:

integration_id avatar_url bio email, phone, SSN in messages

✓ ENABLE_DATA_ANONYMIZATION=true default · deterministic SHA-256 · faculty re-identifies locally from the Canvas roster

The honest answer depends on what you ask the AI to do. There are two zones. Zone one, on the left. Authoring. Pages, modules, syllabus, announcements. When I ask the AI to build a Canvas page from outcomes, roll due dates, or write the Friday lab announcement, the AI never sees a student name. Nothing to leak. FERPA is not in scope for this kind of work. Zone two, on the right. Student data. Grading, summarizing submissions, analytics. This is the harder case, because PII does enter the loop. This is where canvas-mcp does the work. Vishal Sachdev built source-level anonymization into canvas-mcp. The function is anonymize underscore response underscore data. It runs on every API call to a student-data endpoint. Before the response reaches the model, names get hashed. Student underscore eight hex characters. Emails, login IDs, SIS IDs, integration IDs, avatar URLs, bios, time zones, locales, all redacted. The Canvas numeric ID is preserved because canvas-mcp needs it for follow-up operations, but it is not a name a person can read. On the left, what Canvas sends. Maria Rivera, mrivera at usouthal, login mrivera forty-two, SIS J zero zero one two three four five six. On the right, what Claude receives. Student underscore five nine nine four four seven one a. Same record. The work is preserved. The identifier is gone. So I can ask Claude to grade twenty-two submissions and it will evaluate the work without knowing whose work it is. The hash is deterministic SHA two-five-six. The real Canvas numeric ID is preserved alongside the hashed name. When I am done, I take the Canvas roster, hash each student ID myself, and match Student underscore five nine nine four four seven one a back to Maria for grade entry. Re-identification happens on my machine. The model never sees the link. The flag is ENABLE underscore DATA underscore ANONYMIZATION. The source defaults to true. Faculty stays the school official. The AI is the tool. canvas-mcp keeps the boundary.

Accessibility · the 2027 recipe

                    DOJ Title II · WCAG 2.1 AA · Apr 26, 2027 · WebAIM Million 2026
                

One skill file. Four rules. Any course.

FERPA was the first compliance question. Accessibility is the second. Write the rules once. The AI applies them across every page.

FERPA was the first compliance question. This is the second. April twenty-six, twenty twenty-seven. The Department of Justice Title Two deadline for public university web content. Mobile is well over the fifty thousand population threshold the rule uses, so the University of South Alabama is in scope. Every Canvas page in every course we teach has to meet WCAG two point one Level AA by that date. The answer is on the left. One markdown file. About eighty lines. Lives in your project's dot claude slash skills folder. The AI loads it whenever you touch a Canvas page and applies every rule in it. Section one. Color contrast. WCAG one point four point three. Eighty-four percent of pages on the public web fail this. For every text node, compute contrast against the background. If less than four-point-five to one, propose a darker color. GATE per fix. Section two. Alt text. WCAG one point one point one. Fifty-three percent fail. Walk every image and PDF, draft alt from the caption and surrounding context, flag any scanned PDF that needs OCR. GATE per image, because alt text is judgment-heavy. The AI gets it close. You decide whether close is right. Section three. Headings and link text. WCAG one point three point one for headings, two point four point four for links. Forty-six percent of the web fails empty-link text, the click-here problem. Keep heading order sequential, no jumps from h2 to h4. Rewrite every vague link to describe where it goes. GATE per diff. Section four. The legal trap, worth pausing on. The Title Two rule has a third-party content exception. Faculty read it and assume it covers anything they did not author. It does not. If the syllabus requires a publisher portal, an embedded simulation, a third-party video, those are inside the scope. You picked them. The rule lists every external tool the syllabus requires, finds its accessibility statement, and FLAGS any that fail. The GATE here is whether to keep using the tool at all. On the right of the slide you see what each rule produces. Gray body text turning dark. An empty alt becoming a real description. A skipped heading becoming sequential. An unaudited publisher link becoming one with a documented status. Same shape as everywhere else in this talk. The AI proposes. The GATE pauses. You decide. The compliance work that used to be a once-a-year audit project becomes an ongoing pass that runs whenever you touch a page.

Try it yourself · resources + acquisition order

                    All open source · all faculty-built or open standard
                

Where each component lives. What you read or install.

01 The runtime · Component 1

Claude Code.

Anthropic's agentic CLI. Reads your files, calls tools, waits at every gate. The same agent you saw in the live demo.

Where claude.com/claude-code

02 The bridge · Component 2

canvas-mcp.

The faculty-built bridge. Speaks the open Model Context Protocol. Mint your Canvas token in Account → Settings.

Where github.com/vishalsachdev/canvas-mcp

03 The discipline · Component 3

Skills, hooks, gates.

Markdown rules + automatic checks + the four-letter yes/no menu. None of it is code. All of it is writing. Start with one skill file under .claude/skills/.

Where skillsmp.com

+ Helpful environment

Visual Studio Code.

A free editor for the skill files, hook scripts, and Canvas-content HTML. Claude Code runs in its built-in terminal, so the agent and your files share one window.

Where code.visualstudio.com

Before I close, one slide of resources. Where each piece lives, in the order you would acquire it. First. The runtime. Claude Code. The same agent you watched in the live demo. Lives at claude dot com slash claude-code. Install instructions for Mac, Windows, Linux right on the page. This is the only piece that is genuinely technical. Half an hour with a tech-friendly colleague, a graduate assistant, or an Innovation in Learning Center consultant, and you are through it. After that, the terminal is yours. Second. The bridge. canvas-mcp. The faculty-built Python project that connects Claude Code to Canvas. Lives at github dot com slash vishalsachdev slash canvas-mcp. Free, open source, MIT license. The same helper installs it in the same sitting. The only thing you provide is your own Canvas access token, which you mint yourself in USAonline account settings. Two clicks. No IT ticket. The FERPA protections I just described all live in this layer, default-on. Third. The discipline. Skills, hooks, and the GATE. The part that is yours, and the part that nobody else can write for you. A skill is a markdown file with your course rules in plain English, saved under dot claude slash skills slash. Hooks are tiny shell commands that fire automatically. The GATE is the four-letter yes/no menu you saw on every demo today, C R S N. None of this is code. All of it is writing. If you can write your course rules clearly enough to hand to a substitute, you can write them clearly enough to hand to the agent. The community library at skillsmp dot com already has skills other faculty have published, accessibility, grading, document conversion, ready to read and adapt. Plus one. The environment. Visual Studio Code. A free editor where your skill files, your hooks, your Canvas-content HTML, and Claude Code itself all share one window. Not required. The terminal alone will run it. But it is what I use, and it is what your tech helper will probably suggest. That is the road. Three things to acquire in that order, plus one editor to live in while you build. The agent proposes. You decide at every gate. The deciding stays yours regardless of how far you go.

Caution · the same API works both ways

                    sources verified: Instructure docs · Oct 2025 product blog
                

Everything I just showed you, any student in your course can do too.

The Canvas REST API does not distinguish a faculty token from a student token.

The reality

The API is symmetric.

A student can mint a personal access token in under a minute. The same endpoints I demoed expose their own quiz attempts. List the quiz. Read the questions and the answer choices. Submit. Instructure publicly confirmed in October 2025:

“students setting up AI integrations to automatically review and complete assignments on their behalf.”

community.canvaslms.com / strengthening-security blog (Oct 2025)

The answer

Design beats detection.

Proctoring tools cannot see an API client running off-machine. The durable fix is two-layered. At the institution, admins can disable student access tokens with one permission flip. At the course, the answer is assessment design: in-person proctored fact recall, frequent low-stakes formative work, project-based application, oral defense.

Stanford Teaching Commons · MIT TLL · AAC&U

Before I let you go, one slide I would not have shown a year ago, but I have to show now. Everything I just showed you, any student in your course can do too. The Canvas REST API does not distinguish a faculty token from a student token. The same workflow I used to push pages can pull a student's own quiz questions and submit answers. Pretending that is not the case is the bigger risk than naming it. On the left, the reality. A student goes to Account, Settings, Approved Integrations, and mints a personal access token in under a minute. The same endpoints I demoed list quizzes, retrieve the questions and answer choices for that student's own in-progress attempt, and submit answers. Instructure publicly admitted this in October twenty twenty-five, in a product blog announcing tighter token controls. The quote, students setting up AI integrations to automatically review and complete assignments on their behalf. That is the platform vendor naming the problem. On the right, the answer. Notice what is not on this slide. LockDown Browser. Camera proctoring. Honor-code pledges. None of those see an API client running on a phone in the student's pocket. The durable fix is two layers. At the institution, a Canvas admin can disable student personal access tokens with one permission flip. One checkbox. Most institutions have not flipped it. I have flagged this in writing to our Innovation in Learning Center. At the course, the answer is assessment design. In-person proctored fact recall. Frequent low-stakes formative checkpoints. Project-based application. Oral defense. Detection is an arms race. Assessment design is not. The workflow I showed you is symmetrical. The answer to that symmetry is how you ask students to demonstrate learning, not what software you license to watch them.

Thank you.

                    Questions?
                

                    scan to revisit the deck
                

aud613.niyaziarslan.com/presentations/cotl-2026.html

Disclaimer: independent academic work. Not sponsored by or affiliated with any company mentioned in this presentation.

Niyazi Arslan · University of South Alabama

niyaziarslan@southalabama.edu

Faculty-Directed AI for Canvas Course Development

Half of US higher-ed enrollment runs on Canvas.

Building a full Canvas course.

Hidden behind every stage.

How can AI help with each one?

Where AI fits.

What most faculty do today.

AI inside Canvas, systematically.

Claude Code.

MCP. The Model Context Protocol.

canvas-mcp + the Canvas API.

Skills, hooks, gates.

The full loop.

What you can do on Canvas with this.

From your phone, by voice.

One real prompt. One real Canvas page.

AUD613 + AUD647 · Spring 2026.

Two zones. Two answers.

One skill file. Four rules. Any course.

Where each component lives. What you read or install.

Everything I just showed you, any student in your course can do too.

Thank you.

Faculty-Directed AI
for Canvas Course Development