This post was contributed by Ivar Anderson, student and instructional aide in the (fabulous and wonderful) Electromechanical Technology (EM-Tech) program at Columbia Gorge Community College.
STEM study guides and textbooks run on images. Circuit diagrams, hydraulic schematics, and practice problem figures aren’t decorative; they carry the instruction. For a blind or visually impaired student using a screen reader, a missing description or one that identifies a circuit diagram as “a circuit with components” fails them the same way a blurry photocopy would fail a sighted student. But creating useful descriptions of technical figures is no easy feat: they tend to be longer, more precise, and more demanding to produce than most alt text. They also require someone who understands the content well enough to know what a student needs from the image in order to do the work.
Here, we see a bottleneck. Manually producing accurate figure descriptions for STEM course materials requires not only subject matter expertise, but time. And lots of it.
AI assistance might help carry the load, but only if it is guided by domain-specific reasoning rather than general-purpose description logic. A tool that doesn’t know the difference between a component label and a functionally critical detail will produce descriptions that are present but not useful. I developed a Figure Description Skills Package to encode domain-specific reasoning so that genuinely useful figure descriptions could be generated for open education resources, rapidly and reliably.
The workflow is built around .docx files but the description-writing skills work regardless of format or document type. Study guides were the starting point, but the package handles textbooks and other technical instructional materials as well.
The package, a single SKILL.md file, is publicly available under an MIT license, and can be adapted for other disciplines, formats, or institutional workflows. It can be found at Circuit Description Generator on GitHub. You can find my contact information on the GitHub profile page; feel free to reach out.
What the Package Does
The package is a set of four integrated instruction files, called skills, designed to be loaded into an AI assistant session (developed in Claude Sonnet 4.6; likely compatible with GPT 4.5 or later) before beginning any figure description work. Together they define a complete, reproducible workflow: examining images carefully before writing, producing descriptions calibrated for screen reader delivery, inserting those descriptions into Word documents correctly, and handling the domain-specific requirements of technical schematics.
The skills were developed and tested against real study guides in electrical and fluid power technology programs. They reflect iteration with a specific set of failure modes: AI systems that apply system-level labels without verifying that the underlying connections are actually shown, that omit functionally critical details, that produce descriptions too generic to distinguish between otherwise similar figures, or that inventory components without tracing the circuit path a student needs to reason about.
The Four Skills
Skill 1 — Figure Descriptions for Study Guides establishes the core content and formatting rules. It specifies how descriptions should be labeled, how length should be calibrated to the complexity of the image, and, critically, what to include or withhold depending on what the paired student instruction is asking the student to do. A figure paired with “calculate the output voltage” should include all given values but not the answer. A figure paired with “explain why current flows in this direction” should include the full circuit topology. Getting this distinction right is what separates a description that is genuinely useful from one that is merely present.
Skill 2 — Figure Description Docx Workflow covers the mechanics of inserting descriptions into Word documents: determining the correct image order from the document’s underlying XML (which does not match filename order), unpacking and repacking the file without corrupting it, and placing description paragraphs in the correct position with correct formatting. It also documents known issues: pre-existing validation errors in some source files, images with alpha channels that render as black on white in some environments, and cases where image values are ambiguous enough to require verification before writing. The docx format is where this workflow was built and tested, but the core logic — image inventory, order determination, description placement — transfers to other document formats. Institutions working in LaTeX, HTML, or other structured formats will find the approach adaptable even where the specific file-handling steps differ.
Skill 3 — Hydraulic and Pneumatic Schematics supplements Skill 1 with rules specific to fluid power circuit images. These require different handling from electrical circuits: the relevant question is usually about flow path and component behavior rather than voltage and current, and the connection topology must be traced completely rather than described symbol by symbol. This skill also specifies how to identify which student instruction is paired with a given image when the document layout makes this ambiguous. For study guides in this domain, I also provided schematic symbol references to Claude in a project file.
Skill 4 — Circuit Analysis addresses electronic circuits, amplifiers, and component-level schematics. It establishes a verification workflow and documents the most common failure modes in AI-generated circuit descriptions. This skill is more specialized than the others and will be relevant primarily for electronics programs.
