Back to portfolio

AI Agents workflow for course production

Created Agent Skills in Claude Cowork to support Canvas LMS course production across all online programmes, with instructions tailored to each programme’s requirements.

At a glance

Context
Canvas LMS course production
Role
AI Enablement
Timeline
August 2026
Tools & Skills
Claude Cowork Claude Code Codex Agent Skills Canvas LMS Google Workspace GitHub Wistia BugHerd H5P HTML/CSS/JavaScript

Project overview

I created AI agents using Agent Skills in Claude Cowork to support Canvas course production across different educational programmes. Each agent had a defined task, with instructions tailored to the programme’s content, branding, and build standards.

The objective was to reduce repetitive production work and support consistent programme set ups, course builds, feedback implementations and reviews creating more time for content decisions and testing the learner experience.

The challenge

A request to build two programmes in parallel, comprising 10 Canvas courses within four months, triggered the work. Setup, content builds, media placement, and review all needed to progress within the same delivery window. Rather than create a workflow only for those programmes, I developed reusable Agent Skills that could be adapted to the requirements of other online programmes.

My approach

I developed the Agent Skills from real course-production tasks. I first completed each task in Claude Cowork, checked the results, and turned the effective process into reusable instructions tailored to each programme’s requirements.

  1. Sub-account set up

    I used Claude Code and Codex to improve the global CSS, JavaScript, and reusable HTML markup, then created an Agent Skill to review the Canvas setup against the specific educational programme’s requirements.

  2. Storyboard to Canvas build

    • I created an Agent Skill to retrieve approved storyboards from Google Workspace and convert their content into programme-specific HTML in Canvas. It adds images and document downloads from Canvas Files, creates glossary term modals, and embeds videos, H5P activities, and GitHub Pages content.
    • The skill populates the matching module, creating pages, assessments, quizzes or discussions as needed. It presents a dry run before making changes and reloads the pages to verify the build.
  3. Video placement

    • Course videos are hosted in Wistia and listed in a Google Sheet separate from the storyboards. Manually finding each video and adding its embed code to the Canvas markup is time-consuming.
    • I created an Agent Skill to use the sheet to locate each video in Wistia, retrieve its embed code and transcript, and place it in the correct Canvas video markup.
  4. Feedback implementation

    • QA feedback is recorded in BugHerd as tickets linked to course content. I created an Agent Skill to collect all outstanding tickets assigned to me, check them against the approved storyboard, and separate actionable corrections from items needing clarification or blocked by missing information.
    • It presents a dry run showing the current and proposed Canvas content before applying approved changes. It then verifies each update, confirms the page’s publication status is unchanged, and returns the actioned ticket numbers for me to move to review in BugHerd.
  5. Build review

    • I created an Agent Skill to compare completed Canvas pages with approved storyboards and build standards before publishing.
    • It checks content accuracy and page order, components and media, accessibility details such as alt text and transcripts, and course navigation and completion requirements.
    • It also tests hyperlinks and checks that internal course links remain valid when cloned.
  6. Course cloning and verification

    • I created a plugin containing three programme-specific Agent Skills to prepare existing courses for a new student cohort.
    • The plugin clones the course, applies course settings, adds teachers, connects the appropriate BugHerd board, and configures the student-support link and module locks. It asks for confirmation before creating anything and verifies the setup afterwards.

Adapting the agents to each educational programme.

I tailored the agents to each programme’s approved source materials, branding, markup, and build requirements. This made the workflow reusable across online programmes without treating their content or standards as interchangeable.

The impact

The workflow reduced production time and rework, creating more capacity for content decisions and deeper quality checks across the course-production process.

Three improvements followed:

  • Time savings across production tasks

    Module production fell from approximately one working day to one hour, video placement from three hours to 20 minutes, and build-review time by 50%. The workflow addressed multiple stages of production rather than speeding up a single task.

  • More capacity for people, less repetitive work

    Faster production and fewer correction cycles gave learning designers more time to focus on content. BugHerd tickets reduced by 50%, reducing the work needed to document, implement, and recheck issues across the production and QA teams.

  • More thorough checks of the learner experience

    The time recovered created space for deeper QA, not just faster delivery. I could investigate content, accessibility, navigation, and functionality in more detail, identifying issues that might previously have been missed under time pressure.

Reflection

“The extra effort I put into creating end-to-end agents paid off by shifting my role from doer to reviewer. I could focus more on judgement and quality, while reducing repetitive work and rework across the production process. The benefit extended beyond my own workload, giving colleagues more time to focus on theirs.”

— Aida

Get in touch

Like this kind of work?

Happy to chat about full-time roles or projects — pick whichever option works best for you.