I’ve been teaching people how to use AI effectively for a while now. One of the things I keep coming back to is the difference between using AI and delegating to AI. Last week I had a chance to put that distinction into sharp relief.
The Task
I had a completed book—InDesign Masterclass: Text Techniques, 171 pages—and I needed to convert it into a structured online course for a learning management system. That means taking the book’s content and breaking it down into chapter folders, individual lesson files, extracted images, and properly formatted Markdown—one file per lesson, ready to copy and paste directly into the LMS.
Doing that manually would mean opening the InDesign files that comprise the final book, identifying every chapter and lesson heading, copying the text for each lesson, reformatting everything (bold for UI element names, code formatting for code, arrow notation for menu paths), saving each lesson as its own file, extracting and organizing hundreds of images, and numbering everything in order.
Multiply that by 141 lessons across 7 chapters and you’re looking at days of tedious, error-prone work.
What I Did Instead
I described the job to an AI agent.
Not “summarize this PDF.” Not “give me a list of the headings.” I described the full production task: read the PDF’s heading structure, use H1 tags to define chapter folders and H2 tags to define individual lesson files, extract and organize every image, apply semantic formatting rules throughout the text, and output everything as numbered folders of numbered Markdown files.
The agent—Claude Cowork—wrote the extraction script, ran it, and started debugging.
Where It Got Interesting
This is the part that doesn’t show up in demos: real documents are messy.
The PDF had quirks that I never would have caught manually—and honestly, wouldn’t have thought to look for. Soft hyphens embedded in the text were splitting words across PDF text blocks, producing artifacts like “con trols” and “parenthe sis” in the output. Page range calculations were inverting for back-to-back lessons on shared pages. Bold formatting patterns were overlapping and producing ****double-wrapped**** output.
I didn’t diagnose any of those. The agent did. I course-corrected direction a few times—images should be in a subfolder per chapter, number the lesson files, not just the chapters—and it adapted. Each fix was surgical.
The final output: 141 lesson files across 7 numbered chapters, 552 extracted images (with their Alt Text preserved from the PDF), semantic formatting applied throughout.
What This Actually Demonstrates
This is what agentic AI looks like when it’s working—not a chatbot answering questions, but an agent executing a production workflow: writing code, running it, reading the output, identifying failure modes, fixing them, and delivering a finished artifact.
The key shift is in how you frame the work. Instead of asking what should I do?, you define the outcome and let the agent figure out the path. That requires being precise about what “done” looks like—the folder structure, the file naming convention, the formatting rules, the edge cases. The more clearly you can describe the deliverable, the more the agent can operate without hand-holding.
That framing is a skill. It’s learnable. And it’s exactly what I teach.
The Leverage Is Real
If your team is still using AI as a fancy search engine—asking it questions, getting answers, doing the work yourself—there’s a significant amount of leverage you’re leaving on the table.
The question worth asking isn’t “can AI help me with this?” It’s “can I describe this well enough that AI can do this for me?”
In most cases, if you can describe it, it can do it.
Pariah Burke teaches AI productivity and workflow automation, along with InDesign, Illustrator, and publishing tools. Learn more at iampariah.com.